"I like China… People say, oh, you don't like China, no, I love them."

- Donald J. Trump

Excited to return to writing my updates after a brief, unannounced hiatus. This year's central theme appears to be China—in more dimensions than you might expect. Whether it's AI, biotech, academia, or tariffs, China is squarely in the spotlight. Even in defense, where 8VC has built a robust investment franchise, China remains at the forefront of everyone’s attention.

It’s not widely known, but I've been closely studying China since my days at Clarium, beginning in 2008. During that period, I immersed myself deeply in Chinese history, politics, and macroeconomic analysis, as I primarily covered the region for the fund. It wasn't unusual for me to stay awake until 1–2 AM trading Chinese markets, only to return to the desk in time for the 6:30 AM market open the next day!

It was a rewarding experience, and I've maintained my deep interest in the region ever since—leaving me reasonably well-qualified to comment on the current political and economic landscape. At that time, my view was that China would continue industrializing and emerge as a leading global power. Internally, this perspective was somewhat controversial due to the numerous risks facing the country, particularly the massive property bubble and rampant corruption at local government levels. Many believed such growth was unsustainable. However, I thought otherwise, primarily because of the enormous productivity gains from urbanization. When a Chinese farmer moved from a subsistence-level plot to factory employment, their productivity increased tenfold. Even if half of that gain was wasted, the net increase in prosperity would still be substantial.

China has come a long way, transforming from the impoverished nation it was in 2008 into a global powerhouse. Back then, we used to say China advanced primarily by copying others—but today, China leads in many industries. A recent study by the Australian Strategic Policy Institute revealed that China now dominates global research in 57 out of 64 key technological fields, up from just three fields in 2007. Anyone working closely with China knows it's no longer merely a low-cost outsourcing destination; it now produces high-quality products across numerous sectors. Even in academia, Chinese research has surpassed the U.S. in terms of top citations.

And there's the rub: China initially competed through copying and intellectual property theft but has now, in some areas, surpassed the West in critical research. As China emerges as a potential peer adversary, its unique rise is causing concern in the corridors of power, prompting reactions from both government and industry. I'm not an extreme China hawk, but I do believe action is necessary given the trajectory we're observing—particularly regarding the intellectual property we've already lost.

Without further pontification, let me dive into what I believe this means for the economy and startups more broadly.

Macro Update & Tariffs - China Eats your Portfolio

In my view, the first major consequence of China’s rise manifests itself in our current trade debate. Driven by substantial trade deficits—most notably with China—policy tensions escalated dramatically on April 2nd when President Trump announced his "Liberation Day" tariffs on all countries, subsequently liberating Americans from about 15% of their wealth. Commentators debated whether Trump intended to free Americans from dependence on foreign nations or, in a more Buddhist sense, liberate us entirely from our material desires.

The administration has provided very little clarity regarding the intent of these tariffs. Ostensibly, they aim to reshore key industries such as energy, pharmaceuticals, and semiconductors—yet they've exempted precisely these sectors from the "liberation" tariffs. Statements from administration officials have been equally lacking in logical consistency. So far, I've heard three incompatible short-term outcomes promised (though the long-term message is clear: a golden age awaits):

1. There's nothing to worry about. Wall Street's rush to project lower growth is excessive. Everything will be fine in both the short and long term as we reshore industries.

2. It will be tough in the short term, but worth it. The tariffs will remain, eventually leading to an American economic boom as companies reshore.

3. We have more than 50 countries negotiating, and we'll ultimately reach a world of uniformly lower tariffs.

That leads me to a few conclusions with relatively high confidence. First, there is no master plan—the administration is simply reacting on the fly based on the circumstances presented to them. Whatever deal ultimately emerges will retroactively become the "plan." This explains erratic moves, such as implementing a 90-day tariff pause when markets crash, or constantly shifting exemptions. Second, anyone claiming to know exactly what will happen next is lying. We might soon witness a complete cessation of trade between the U.S. and China—or, alternatively, a swift deal could be reached. Similarly, global barriers might be reduced if countries cooperate to isolate China —or tariffs could persist indefinitely if they're deemed beneficial. Trust no analysis; the only prudent course is patience.

On interest rates, the administration's messaging is equally inconsistent. Trump clearly wants lower rates, yet both his tariff policies and his goal of reducing the trade deficit are likely to push rates higher. The tariffs themselves have a straightforward effect: they'll drive inflation upward—as Bank of America recently projected for year-end—and rising inflation inevitably leads to higher interest rates.

Reducing the trade deficit will also push interest rates higher, though the mechanism is less intuitive. The reason lies in the balance of payments. You've likely heard the claim that "China (or foreigners generally) finances our fiscal deficit." This is true due to the nature of the balance of payments, which must always balance: every dollar entering the U.S. must either purchase a good or an asset. I recently explained this on Twitter.

Without diving too deeply into economics, consider the current account (mostly the trade deficit) and the financial account (assets bought by foreigners). If the trade deficit shrinks, foreign purchases of U.S. assets decline correspondingly. This forces domestic investors to absorb more assets, driving interest rates higher (and prices lower).

Amid uncertainty, trade tensions, and rising volatility, the U.S. is trading like an emerging market. Typically, crises push stocks down and drive investors toward bonds and the dollar for safety. But with confidence in the dollar shaken and volatility driven by U.S. policy itself, we’re seeing stocks, bonds, and the dollar all falling simultaneously—reflecting capital flight rather than a flight to safety.

The conclusion of my macro update is that we genuinely don't know much. I'm confident inflation will stay elevated for now, and there's certainly recession risk. However, exactly where we'll land remains uncertain. If you're running a company, enjoy the ride!

ASIDE: I'm one of the few proponents you'll meet who supports targeted tariffs, but my frustration with the administration lies in their poor execution. A current account deficit is acceptable—we benefit from inexpensive foreign capital inflows. However, offshoring critical industries creates national security risks, and foreign trade barriers disadvantage American businesses. An effective policy would (1) identify and privately negotiate the removal of unfair barriers, using tariffs as leverage if needed; (2) strategically impose tariffs in sectors we want to reshore; and (3) implement these policies permanently, clearly, and with incentives, providing U.S. companies sufficient capital, certainty, and time to adapt. The administration falls short precisely on this third point, causing uncertainty and unnecessary financial burdens.

China Eats AI

China shocked the world in January with its release of R1, causing widespread panic. I found these reactions misguided (I'll aim to comment more promptly next time). Initially, there were fears that R1—open-source and near-frontier in capability—would commoditize AI, echoing the common adage that “when China enters an area, profits disappear.” But AI has been steadily commoditizing and becoming cheaper from the start; R1 hasn't materially altered that trajectory.

There was also concern about R1 being trained for only $6 million. However, this isn't the first time a lab has achieved near-frontier performance with less than $10 million. As I mentioned previously, Reka accomplished something similar—their core model reached #7 on LMSys. It's important to note that this $6 million figure covers only the direct cost of training, assuming the team already knows exactly how and what to train (as Reka did). It doesn't include the research costs needed to reach that point, which DeepSeek has not disclosed.

There was also considerable speculation claiming R1 had trained on outputs stolen from OpenAI. I'm skeptical of these claims, especially given DeepSeek's consistent track record of publishing high-quality research. It's not surprising that a team producing strong research would also deliver strong models; there's no reason to diminish their achievements. In fact, they introduced several impressive innovations, notably in wide MoE architectures, multi-latent attention, and low-precision training.

Ultimately, it wasn't the lower cost of training or inference that made R1 impressive. Instead, what genuinely surprised people was China's ability to release a near-frontier model—ranking #2 on LMSys at launch—despite extensive export restrictions. Yet even this shouldn't have been entirely unexpected; Alibaba’s Qwen models have consistently demonstrated strength, though R1 clearly exceeded them.

This fear has prompted the administration to enact further restrictions on GPU sales, including a ban on H20 exports and potential restrictions on DeepSeek—which, in my view, would violate the First Amendment.

The LLM model tier isn't the only area where China is competitive.

1) A Chinese company recently released a general AI agent called Manus, capable of using tools such as web browsers to perform various tasks. It's also notably proficient at code generation, and early reviews on Twitter have been quite positive. I haven't personally tested it yet—they're severely compute-constrained—but it's reportedly impressive. Benchmark recently invested at a $500M valuation [aside: I thought we weren't allowed to invest in Chinese companies anymore?].

2) In response to export restrictions, Huawei has made impressive strides in AI silicon development, recently introducing their own rack-scale AI architecture, the CloudMatrix 384, which directly competes with the GB200 NVL72. SemiAnalysis provided an excellent breakdown here. The TL;DR is that it's a strong product—less power-efficient than the GB200 but superior in some respects. While software gaps remain, especially regarding CUDA compatibility, it’s clear that China is actively advancing. This represents a significant threat to NVIDIA, considering China historically accounts for 20-25% of their revenue.

3) Alibaba released their frontier video-generation model, Wan 2.1. It's extremely impressive—on some benchmarks, it achieves frontier-level performance even compared to closed-source alternatives. Take a look at these examples here!

Are Chinese models near-frontier? In many cases, yes—most people likely can't distinguish state-of-the-art Chinese open models from leading closed-source US models. With Huawei’s CloudMatrix 384 release, a nationwide push to replicate the CUDA ecosystem is inevitable. Although this will take years, once China achieves it—and once SMIC can manufacture HBM—they'll have fully decoupled from the US in AI capabilities. This is why I believe NVIDIA’s GPU restrictions were premature. Limiting exports is a strategic move we can only use once, prompting China to rebuild their hardware and software stack. The optimal timing wasn't now, when AI's economic impact is still nascent, but rather in 3–4 years, when losing access to CUDA would truly hinder their progress.

So, will China ultimately win? No. GPU restrictions may not significantly impede their model training, but they will severely limit inference capabilities. We're already seeing this with DeepSeek and Manus—strong models with the applications made nearly unusable by compute constraints. As a result, China will likely focus on open-sourcing AI models, allowing others to provide inference hardware. This accelerates AI commoditization—a trend I've highlighted for years—standalone models remain weak businesses. Real value lies in the software and products built around these models, and historically, China hasn't excelled at application software. Build product!

China Eats Bio

In my view, the most concerning and immediate threat to America’s innovation economy comes from China’s activities in biotech. In this high-tech sector, China's competitive advantages and regulatory structure could effectively commoditize innovation. This presents a major challenge to U.S. biotech innovation and may pave the way for significant competition against U.S. pharmaceutical companies in the future.

Over the past decade, China has developed world-class capabilities in two critical areas of therapeutic development: medicinal chemistry and biologics discovery/manufacturing. In specialized niches such as bispecific antibodies, Chinese companies have become genuine global leaders—WuXi XDC, for example, is now the premier player in bispecific development.

But China hasn't limited itself to just providing services. Chinese companies now actively conduct clinical trials using infrastructure that's both rapid and cost-effective. They can run trials using non-GMP clinical materials, streamlined IRB approvals, expedited regulatory timelines, and recruitment from a significantly larger population. As of 2023, China accounted for 28% of global clinical trial initiations—a figure that's continuing to rise.

Now, more novel molecular entities originate from China than from Europe!

Now, here's the rub: The Chinese excel at rapidly and cost-effectively creating molecules and running clinical trials. They've combined these strengths with a fast-follower strategy in drug development. When a therapeutic target gains popularity in the U.S., often via biotech development, Chinese companies swiftly replicate it and advance it through Phase 1 trials—often faster than their U.S. counterparts. Afterwards, they typically seek to out-license these molecules to U.S. pharmaceutical companies at significantly lower prices.

This fast-follower strategy has enabled Chinese molecules to capture a substantial share of licensing activity. In 2024, molecules originating from China accounted for 28% of all licensing deals—a figure that has risen to 33% in 2025. Furthermore, nearly 30% of major pharmaceutical licensing deals (those exceeding $50 million upfront) involved Chinese biopharma companies in 2024, a notable increase from 20% in 2023 and virtually none five years ago!

These molecules aren't inferior or mere copycats—many represent genuine frontier innovations. For example, Carvykti, a BCMA-targeted cell therapy for multiple myeloma (MM), was licensed by J&J through a co-development deal with China's Legend Biotech and generated nearly $1 billion in sales for Legend in 2024. Fundamentally, Carvykti is Legend's drug, commercialized in partnership with J&J.

The surge in Chinese biotech activity poses an existential threat to the U.S. biotech industry. Chinese companies rapidly produce inexpensive fast-follower molecules, complete with Phase 1 data, and license them to U.S. pharma at a fraction of the cost U.S. biotechs typically demand. Given the high expenses of U.S. biotech research, this undercuts the value of original innovations. If this continues, it could severely impact biotech valuations, further challenging an industry where attractive returns are already scarce.

This doesn't apply equally across all areas of biotech. In fields with strong IP protection past composition of matter (bispecifics, TCRm and engineered TCRs), Chinese companies can't easily enter the U.S. market—but in areas without robust IP coverage, significant challenges emerge.

China Eats Defense & Industry

A silver lining of China's rise is that the potential threat of conflict in the South China Sea or elsewhere has spurred renewed innovation and readiness within U.S. defense, as I noted previously. The ongoing war in Ukraine has already highlighted critical vulnerabilities in America's defense supply chain—particularly regarding preparedness for a potential conflict with China, which we all hope never occurs. A recent report presented to Congress by the Commission on the National Defense Strategy concluded that "unclassified public wargames suggest that, in a conflict with China, the United States would largely exhaust its munitions inventories in as few as three to four weeks, with some crucial munitions (e.g., anti-ship missiles) lasting only a few days. Once expended, replacing these munitions would take years."

Among the commission's recommendations are proactive engagement with the private sector—particularly startups focused on emerging information technologies—to build a modernized industrial base. They also recommend reevaluating counterproductive regulations that currently hinder defense technology procurement and sales.

This geopolitical push, driven implicitly by China and Russia, has served as both a wake-up call and a boon to startups focused on defense and resilience.

Across the defense ecosystem, significant initiatives are underway to enhance capabilities on land, sea, and air, and to expand industrial capacity. Our company, Saronic, recently announced the Gulf Craft shipyard in Louisiana, which will build our new 150-foot Autonomous Surface Vessel (ASV), Marauder. We're bringing shipbuilding back to America.

Similarly, Anduril announced the opening of its Arsenal-1 hyperscale manufacturing facility in Ohio, designed to produce tens of thousands of autonomous weapon systems. This facility joins their solid rocket motor factory in Mississippi, robotic submarine facility in Rhode Island, and launched-effects factory in Georgia.

This aligns closely with lessons from the war in Ukraine, where drones reportedly cause up to 70% of battlefield fatalities, surpassing traditional munitions. Future wars will increasingly rely on drones and autonomous systems, posing significant threats to human combatants. Such a threat is concerning given China's dominance in the drone industry: Chinese companies control about 90% of the U.S. consumer drone market, 70% of the industrial market, and approximately 70% of global civil drone sales as of 2023. Continued reliance on drones in warfare makes this dominance a critical strategic risk.

U.S. companies are actively responding. Anduril has developed a comprehensive counter drone suite, while Epirus is focused on neutralizing drones using long-pulse, high-power microwaves. These companies are part of a rapidly growing ecosystem of businesses preparing to compete in the future of warfare. China's current dominance underscores significant opportunities for new U.S. entrants, particularly in technological innovation and industrial capacity.

Other supply-chain vulnerabilities are also emerging as critical concerns, creating opportunities for U.S. companies to reshore manufacturing. Any strategically important sector dominated by China poses potential risks. For example, during the pandemic, fears arose regarding U.S. access to life-saving medicines due to possible hoarding by countries like China. This concern motivated 8VC to found Resilience Bio, ensuring reliable, friend-shored manufacturing in this vital sector. Similar companies can and should be established in other critical areas.

Chinese involvement in these critical domains, from military threats to dominance in strategic industries, presents significant opportunities for U.S. startups. Countering these threats and reshoring essential industries will be major areas of opportunity over the next decade.

Conclusion

The point of my update isn't to suggest that all hope is lost. In fact, China faces significant long-term structural challenges of its own—inefficient governance, corruption, and demographic decline driven by poor top-down policies. However, the key takeaway is that China is a more formidable competitor than many realize, posing genuine strategic threats to American technological dominance. Moreover, any trade war with China carries the inherent risk that other global players might align themselves with Beijing, given its near-frontier technological ecosystem.

So what's the conclusion for startups? I think it's nuanced. Geopolitics isn't typically viewed as a major driver in private markets, largely due to their long investment horizons. However, I'd summarize the key takeaways as follows:

• For biotech companies: China poses a significant threat. To succeed, stay lean, move quickly, and focus on frontier research protected by strong IP (some small molecules, bispecifics, TCRm, engineered TCRs, etc…).

• For startups unaffected by tariffs: Stay focused and execute. Even during a recession, private markets likely have ample liquidity.

• For startups impacted by tariffs: Adopt a wait-and-see approach—pause fundraising, conserve cash, and anticipate resolution within 8–12 weeks.

• For defense and resiliency startups: BUILD!

My parting message for all Americans is to recognize the tremendous agency we have in countering the competitive threats we face. We don't need to rely solely on the government—except to ensure a level playing field. Rather than complaining that competition is unfair, let's simply get better.

Great things I’ve listened to or read:

The Algorithmic Bridge

Google Is Winning on Every AI Front

(PSA: Many people are interested in this post, so I removed the paywall…

Read more

a year ago · 232 likes · 46 comments · Alberto Romero

Air Street Press

There is no scaling wall: in discussion with Eiso Kant (Poolside)

Over the past few weeks, we’ve been through another round of speculation about scaling laws. This time, it’s not been coming from Gary Marcus, but seemingly from staffers attached to frontier labs…

Listen now

a year ago · 9 likes · Air Street Press and Nathan Benaich

"Only when the tide goes out do you discover who's been swimming naked." 

- Warren Buffet

This quarter, I found myself struggling with inspiration on what to cover. Q3 was busy in many ways, particularly with significant developments on the AI front:

• Poolside Raised $500m for AI-powered software development

• OpenAI raised $6.6b

• SuperSafe intelligence raised $1b for AI research 

• Magic raised $320m for coding models

• Groq raised $640m for AI chips

• WorldLabs raises $230m

No peak here yet!

But outside of AI, venture-wise, it wasn't a blockbuster quarter with most of Q3 coming over the summer season. Deal volume seems slightly down (~10%).

In a shorter than normal update I thought I'd put some quick thoughts to paper ahead of the larger Q4 and prospective 2025 update early-Jan. It's going to be a banger.

*Election update: I aim to keep political commentary in these updates minimal, especially as many policy details are yet to be defined. Overall, I believe Trump’s election could be positive for venture capital and the tech sector, given the likely emphasis on deregulation and a more favorable environment for mergers and acquisitions—both of which will benefit tech. I’m considering buying small-cap stocks (tech and otherwise) to take advantage of this potential shift, in addition to my long. I also bought .

Market Update - Number of Investors & Exits Falling

In my second update of 2023, I discussed the apparent oversupply of venture capitalists in the market. I’m cautiously optimistic to report new data suggesting that we may be nearing a correction on the staffing front. Anecdotally, we’ve seen significant shakeups, such as Index Ventures defenestrating 5 of their 9 SF based partners -- a surprising move coinciding with their $2.3 billion fundraise. This reflects the state of the market: while the number of investors may have peaked, dry-powder remains high!

What hasn’t changed is the sheer amount of dry powder available—it remains at an all-time high. However, in my view, the set of investment opportunities has narrowed significantly since 2021 and 2022, declining by an estimated 60–70%.

Most of the decline in the number of investors is due to a slowdown in small, first-time funds.

All of this unfolds amidst a lack of exit activity that has persisted over the past three years.

This is the core challenge: the industry faces a lack of exits while grappling with the rise of ever-larger funds at the high end of the market. Every meeting I have with LPs now highlights this contradiction in liquidity availability. If IPOs and M&A activity remain scarce, and companies continue to stay private longer, new mechanisms will be needed to provide LPs with liquidity in their venture holdings. We either need to address IPO aversion or adapt the ways VC funds achieve liquidity. Selling positions in portfolio companies on the private market will likely need to become commonplace, potentially leading to the growth of funds specializing in this area. Ironically, the fund secondary market seems to be the only private market untapped by VC mega-funds, which remain focused on traditional lines of business.

It’s remarkable to reflect on the unique moment we’re experiencing in the venture world. The number of individual investors is shrinking, opportunities have dwindled, exits have stalled—yet we have more dry powder than ever, increasingly concentrated among the top firms. What a world.

Venture and Dumb Money

The incongruence described above led me to reflect, during a call with an LP, on the unique challenge that exuberance in venture capital markets poses to investors—a challenge that is nearly unparalleled in any other asset class.

I’ve previously shared a piece of wisdom from an early mentor about capital excess: “The remedy to an excess capital environment is that capital gets destroyed” (h/t Patrick Kenary). In most other financial markets, this excess capital—often allocated by what might be called "dumb money"—can be exploited by others. In public markets, you can short stocks; in private equity, you can flip or recapitalize companies; and in real estate, you can sell properties, short debt, or short CDOs, among other strategies.

In venture capital, however, the asset class offers few avenues to profit from the presence of “dumb money.” Secondaries are difficult to execute, there’s no way to short stocks, and no listed debt exists. Instead, dumb money often becomes the top bidder, setting the price for all prospective investors. While an excess of dumb money in most other asset classes creates opportunities, in venture, it acts as a headwind. I used to refer to these players as “venture terrorists” because they blow up returns for everyone. In the rush to deploy capital, I see this dynamic playing out in many AI markets today.

At 8VC, our remedy has been to return to our roots: investing very early, before product-market fit (PMF), to capitalize on the lower cost of capital our companies enjoy when they reach their inflection points. This is the only strategy we’ve identified that indirectly benefits from the presence of “dumb money.”

OpenAI COGS and the Economics of LLMs

The Information recently published a fascinating article detailing OpenAI's financials, offering insight into the company I've never seen before. It provides a clear look into the finances of the leading AI model provider in today's market.

Since it’s paywalled I won’t share the data here, but a quick Google Image search will bring it up if you’re curious.

What immediately stands out is the staggering loss for a business operating at such a large scale. While substantial R&D and CAPEX investments are happening internally, they don’t tell the whole story. Many of you might recall my earlier point that these models are likely among the most quickly depreciating assets ever, with each remaining at the frontier for only a few months. This reality significantly challenges the unit economics of the business.

With this data, I analyzed what the margins for such a business might reasonably look like. Based on a few assumptions—particularly, that the useful life of a trained model is about 6 months (with training costs fully depreciated in COGS), that 70% of training compute is allocated to production and 30% to R&D (excluded from gross margin), and that data costs are split 50% as one-time expenses (depreciated over 5 years) and 50% as licensing fees — I arrived at a fairly shocking -38% gross margin. Even when considering only the obvious COGS components, such as compute costs for running models, the Microsoft revenue share, and hosting expenses, the gross margin still comes out to just 22%. So, I don’t think I’m too far off in my estimation!

If we evaluate the business in terms of contribution margin—and generously assume that costs like compute for training models are entirely fixed—the contribution margin comes out to 22.5%. This means that, assuming headcount remains unchanged, the company would need to generate an additional $20 billion in revenue just to break even. That’s an astonishing scale!

I suspect my contribution margin estimate might be too high, or the company may plan to expand additional money-burning business lines. Reportedly, losses are expected to increase as revenue scales over the next few years. While the company does incredible work, it is undeniably very costly to operate.

This doesn't look like a software company at the moment!

Model CAPEX with Catastrophic Forgetting

One argument for improving OpenAI’s profitability over time is that, as models stabilize, the yearly compute expenditures required to train and maintain them—currently around 50-75% of revenue—should decrease. This would mean the models could achieve a longer useful life than they currently have, which is just a few months. The company could then transition from an R&D-heavy mode to a more maintenance-focused mode, creating an opportunity to start harvesting profits.

I don’t see many people discussing this, but my intuition is that there are two major hurdles to predicting sustainably lower compute costs—even if model architectures stabilize. These hurdles revolve around keeping knowledge cutoff dates recent and integrating new data. If you look at the LMSYS leaderboards, most models are not up to date, with cutoff dates ranging from late 2023 to early 2024.

This is curious, isn’t it? If keeping knowledge cutoff dates recent were straightforward, most providers would certainly do it. I believe the challenges here stem from issues like catastrophic forgetting and post-training constraints, along with data quality and curation—though the latter is a separate concern.

I don’t hear many researchers discussing catastrophic forgetting, so I suspect it’s either a significant problem or not an issue at all. To understand the concept, consider that during pretraining, the training data order is randomized to reduce instabilities. However, when models are fine-tuned on a subset of data or continuously pretrained (CPT) on subsets, catastrophic forgetting can occur. This happens when the model overfits to the new data and loses knowledge from its original pretraining corpus. I believe this is one reason why knowledge cutoff dates aren’t regularly updated—CPT on just a subset of the training data can degrade overall performance. As a result, full, costly pretraining runs might be required to update the model’s cutoff. While there are techniques to mitigate this issue, I’m uncertain how effective they are at the frontier level.

The second issue I’ve heard is the difficulty of adding new knowledge to a model that has already undergone post-training. Academic attempts to "unlearn" certain post-training steps have largely been unsuccessful. As a result, updating the knowledge cutoff would likely require updating the base model itself, which in turn necessitates redoing the alignment process from scratch. Post-training reportedly involves numerous human-driven steps, making it both time-intensive and expensive.

The TL;DR of the above (and I’m about 80% confident in this, as much of the knowledge on training and alignment is proprietary—I’m piecing this together from expert conversations and inferences) is that keeping knowledge cutoff dates semi-recent will likely require large, ongoing training and alignment runs, even if model architectures, sizes, and datasets stabilize.

If no effective techniques are developed to address these challenges, the compute expenditure for foundation model companies is likely to remain persistently high. Models will continue to depreciate quickly, as updating knowledge cutoff dates periodically to stay current will remain necessary. A tough business indeed!

GPU Price Crash

Eugene Cheah, an entrepreneur known for his work on RWKV, wrote an insightful piece about the H100 (NVIDIA's flagship AI GPU) rental market. He explains that outside the tier-1 cloud providers, H100 rental prices have dropped significantly—from $4.70–$8 per hour in early 2023 to just $1–$2 per hour today, a decline of over 50%. Many tier-1 providers, such as Amazon and CoreWeave, still charge $4–$5 per hour, but these rates don’t reflect the broader market. This represents a stark shift from 2023, when investors were scrambling to secure compute time for deals!

This is particularly noteworthy because it renders the economics of buying and renting GPUs quite unfavorable. With rental prices dropping to $2–$2.85 per hour, the resulting internal rates of return (IRRs) are very low—likely too low to be financeable.

This is an important trend to monitor, as unfinanceable IRRs could lead to declining GPU demand for NVIDIA. has been the biggest beneficiary of the AI wave so far, and as I mentioned in my last update, I believe we’re nearing a peak. I don’t think we’re there yet—there’s still significant optimism surrounding their next generation of Blackwell chips, and I’m aware of a few more megaclusters under construction. However, GPU rental rates could serve as an early warning sign, indicating potential weakness in production AI demand relative to CAPEX supply.

As a reminder, I’m extremely bullish on AI over the next 10 years, but I remain skeptical that near-term developments will meet the current inflated expectations. That said, I will let data guide my perspective. I’ll continue to monitor how AI demand in production evolves and track indicators like and GPU rental rates to quantitatively assess demand.

ASIDE: I need a good time series and datasource for GPU rental costs if anyone can provide it

Conclusion

I plan to write a more comprehensive update in my end-of-year note and 2025 lookahead, where I’ll address a few topics I’ve yet to cover—such as my perspective on different types of market leadership and how startups should evaluate their natural growth rate. Until then, onwards!

Interesting Things to Read/Watch:

• Ilya Sutskever recently told Reuters results from scaling pretraining have plateaued. Regular readers heard it here first!

Investors are now tasked with navigating an environment where nearly every layer mentioned is priced for high value-capture, without many proof points on where value will actually accrue. Revenue is being generated, but it's not sufficient. What's uniquely missing is gross margin and defensibility. Many companies in AI are operating with low gross margins in highly competitive spaces relative to near-term demand. How will this all resolve? I don't think most VCs will admit it (there may be a 'Google' out there!), but they don't know either.

It's phase shifts like these that distinguish good investors from great ones. It's common for an investor's success to be limited to the initial macro-regime in which they performed well. Consider investment management as a large-scale model overfitting exercise. In any sustained bull or bear market, the most convicted investors in that trend will outperform. The large population of investors ensures that some will have conviction in the current regime and outperform. However, like most overfit models, they are unlikely to continue to perform well when variables change. Only the greats can persist through multiple cycles. (This, in my view, is why hedge funds have performed so poorly over the past 2 decades). I hope it goes without saying, but my singular goal is to make the transition.

Software seems cheap

In my last update, I discussed how incumbent ISVs will capture a significant portion of AI value and then subsequently contradicted myself. The reality is likely to be more nuanced: some value will be captured by Devin-like multi-agent systems, while incumbent ISVs integrating AI into their product workflows will capture other portions—especially as the models improve. The exact division of value capture is uncertain and will likely depend on the complexity of the multi-agent systems. However, I believe both approaches will ultimately be successful.

The market, however, seems to price the view that cloud software companies will benefit from AI at nearly zero. I've shown this chart many times, but cloud software revenue multiples are trading close to near-term lows. This stands in contrast to a huge boom in AI stocks like NVIDIA, ASML, PLTR, etc.

I think this is wrong, and it's probably the only part of the public and private market that is mispriced today. If you believe AI models will improve significantly over time, ISVs will capture a lot of value. I think that it's very possible that LLMs will resemble 'query processors' more than true platforms. A query processor itself is useful, but it's the tooling and data gravity around it that are defensible (at the limit, LLMs can be just an API call with relevant context). ISVs that are systems of records (SORs) have that data gravity and own the enterprise workflow screens. They are best positioned to benefit from providing a unified, industry-specific AI outcome, which could lead to the LLM layer becoming commoditized over time.

The market isn't pricing in this data and workflow gravity at all! I think public and private cloud software SaaS companies are very undervalued, and I'd love to invest more in them (I've initiated a position in $WCLD from my PA).

AI, the Capabilities Gap & GPU Oversupply

In the past, I've talked a lot about the AI capability gap, and we see it more and more every day. Things you'd hope would work, like structured 'Text-to-SQL,' don't quite work yet. It's baseline functionality like this that is missing—things often work 'most of the time' but not quite 'all of the time.'

The issue to closing the gap is that it seems we are approaching a ceiling on performance with current architectures. From my perspective, we have three major sources of scaling law improvement left before fundamental architectural changes are needed, but these improvements come with diminishing returns. The areas of improvement are: a) better use of synthetic data in various forms, b) the introduction of multi-modal data in training, and c) scaling of compute (we see people building 5-10x bigger clusters and addressing network congestion limitations). Will these improvements push us past the capability gap? I'm not sure, but I hope so; otherwise, we will need a breakthrough in architectural research and that's not deterministic.

At present, and especially if the above scaling methods don't produce sufficient performance gains, I think we're at a point where there is an oversupply of GPUs in the market. We've already stopped hearing companies complain about GPU access—GPUs are plentiful. This will serve as a good litmus test for AI adoption in the market. If inference workloads don't scale, we'll have seen a lot of CAPEX for demand that doesn't materialize. Famous last words, but I think a) we're close to a peak for NVIDIA ($NVDA) in the near future, and b) we're approaching (maybe within a year?) the trough of disillusionment for AI.

Very weakly supporting my disillusionment call is the fact that AI labs continue to shut down, with the latest news being that Adept was acquihired by Amazon. However, the reason my call isn't imminent (one year out) is that new labs continue to form at a faster pace than they shut down, with H raising $220 million, x.AI raising $6 billion, and famed researcher/pioneer Ilya Sutskever starting Safe Superintelligence Inc. Ilya is a legend, but basing a company on safety seems like losing the prisoner's dilemma from day one. How can a model that is more restrained in terms of power compete with those that are less so?

I should remind non-regular readers that I am extremely bullish on AI. However, I believe that, like most technological cycles, we are currently a bit ahead of ourselves in terms of capability versus potential, and the current exuberance is built on potential. AI will be transformational for society, but perhaps not just yet!

Will the impact be similar to that of cryptocurrencies, where the promise was large but the long-term impact has been shallow? Or will it be more like mobile/internet technologies, which, after an initial period of disillusionment (especially with the internet), experienced a long and consistent slope of enlightenment? I tend to think it's more of the latter.

Macro

I don’t have time for a full macro analysis, but it does seem like inflation is abating in the short term, with the CPI falling to 3% in June compared to the expected 3.1%. Looking at monetary indicators, we are likely heading into a period of monetary contraction in many major markets. In the US, money is expected to get tighter in the second half of the year (money can be disconnected from the fed policy rate). Japan seems poised to tighten soon, and China also appears likely to follow suit. Markets could face challenges in the second half of this year, especially given the equity risk premium, which, as I’ve mentioned before, remains at 20-year lows. Monetary flows are key drivers of asset performance, and equity valuations seem extended amid worsening economic data.

I still believe inflationary pressures will return in 2025, especially given the likely political landscape where both parties are expected to increase spending.

Defense Wave

I haven't written about it before, but I'm very excited about the revolution in defense innovation over the last ten years. We've been pioneers in creating defense investing as a viable venture category, and I'm thrilled to say it's finally having its moment. I've decided it's time to write about what has been working and the changes that have enabled startups to succeed in such a challenging area.

There have been a myriad of exciting updates in the past two years, both within our portfolio and beyond, that have cemented defense tech as one of the most exciting and profitable areas of venture investment. Here are a few great companies in the space I'd like to start by showcasing:

1. Anduril: We participated in the Series A, and since then the company has gone on to raise capital valuing it at over $8bn.  The company has made a ton of progress, but recently they cemented their leadership by being selected to move forward on the Collaborative Combat Aircraft (CCA) program (autonomous unmanned aircraft to augment 5th and 6th generation manned fighter aircraft). This is the first time in recent history a critical aircraft project has gone to a non-Prime!

   

   Anduril is also the youngest company since the Korean War to win a Program of Record with their Autonomous Surveillance Towers (AST) program!

2. Saronic: A company we co-founded at 8VC that designs and manufactures autonomous surface vessels (ASVs) for the Navy. They just raised $175m at a $1 billion valuation.

3. Epirus: A company we co-founded at 8VC that creates long-pulse high-power microwave (HPM) technology for counter-drone and drone swarm protection. They recently raised $200m at a $1.35b valuation and shipped four systems to the Army as part of a $66.1m award.

4. Chaos Industries: A portfolio company founded by several ex-8VC employees came out of stealth raising $70m to build advanced sensing solutions.

5. ShieldAI: A non-portfolio company. Developing AI for unmanned aerial vehicles (UAVs). Won a $198m ceiling contract with the Coast Guard for Maritime Unmanned Aircraft System Services.

The zeitgeist in the defense startup ecosystem today looks drastically different than it did for early pioneers of this space, such as Palantir, SpaceX and Anduril. When we discussed investing in the Anduril Series A at 8VC in 2018, there was significant consternation in the tech ecosystem on doing work in defense. It was around this time that Google famously dumped their Project Maven DOD program due to employee unrest, which gives a sense of the strong sentiment against supporting defense innovation in Silicon Valley at the time.

Since then, we’ve been able to witness firsthand the complete shift in perception of defense tech. We recently threw a Defense Hackathon in LA this year and had a great turnout:

A series of changes in government, commercial industries, and culture has shifted defense to a viable venture investment category. The first and most important event leading to change has been the proven success of the vanguard companies; Palantir, SpaceX, and Anduril did what was seemingly impossible -- getting their foot in the door and disrupting decades of defense prime regulatory capture. All three have all shown that startups can compete to build superior products that are more affordable for the government with a compelling investor return profile. Success does a lot to change minds and shows a viable path forward, drawing much more talent into the area.

As part of forging the way, the defense vanguard helped to drive reform that is enabling the entry of smaller innovative companies to further succeed in defense innovation. Palantir did a good job of recently outlining the waves in their defense last supper article:

The current landscape of innovation bucks the trend that emerged due to the incredible consolidation in the defense industry following the end of the Cold War. The government encouraged this consolidation, which effectively reduced competition and created the modern bureaucratic primes. This consolidation, combined with the regulatory capture of these primes, exploited previous defense acquisition methods (like FAR contracting along with large and long bureaucratic PPBE processes, which exist ironically to ensure tax dollars are being spent efficiently and without corruption) to create a “valley of death” for any company that wanted to compete.

The first major breakthrough in the last 15 years that bucked this trend occurred with SpaceX. NASA made a big bet on SpaceX for the Falcon 9 Development, Commercial Orbital Transportation Services, and Commercial Crew Program, where SpaceX ran circles around incumbent primes like Boeing and ULA on firm fixed-price (FFP) Other Transaction Authority (OTA) contracts. Since then, the success of SpaceX, along with companies like Palantir and Anduril, and concerns from Congress that reform is necessary to access the best technology to deter modern foreign aggressors, have led to procurement and funding reforms that make it easier for new entrants to compete.

The relevant changes and reforms to defense procurement for startups, as I see them, are:

1. The creation and growth of the Defense Innovation Unit (DIU): The DIU was established in 2015 to give meaningful contracts explicitly to commercial and early-stage businesses with the goal of fielding and transitioning tech to DOD partners. They've been a crucial supporter of startups in the ecosystem and Congress recently increased their budget to $1b.

2. The growth of Other Transaction Authority (OTA) contracting rather than FAR contracting: OTA contracting, which innovation partners like DIU heavily rely on, is a very flexible method of acquisition which allows federal agencies to engage in research and development, prototyping, and production activities without adhering to the more rigid regulations and requirements of traditional FAR procurement methods. This flexibility is designed to attract non-traditional defense contractors, foster innovation, and accelerate the acquisition process by reducing bureaucratic hurdles and allowing for more tailored agreements that can accommodate the unique needs and capabilities of private sector partners.

3. A push for Firm Fixed Price (FFP) contracts:  FFP contracts are contracts where the government and contractor agree on a set price for the work to be performed, regardless of the actual costs incurred. This contract type shifts the financial risk of cost overruns to the contractor, incentivizing them to control costs and work efficiently, as opposed to more traditional Cost Plus contracts (costs are covered and a margin is added on top) which incentivize increasing costs to grow profits from a fixed margin. Startups have been more competitive in securing FFP contracts compared to traditional Primes due to their leaner cost structures and ability to rapidly innovate. Certain branches like the Air Force and Space Force have explicitly stated their support of FFP contracts, while others still remain stubborn supporters of Cost Plus. By placing the risk on the company,  FFP contracting has been particularly advantageous for startups, as many Primes have publicly stated they are reluctant or flat out won’t bid on FFP contracts due to the financial risks involved and massive losses some have incurred in recent years from them. The continued push for FFP contracting helps ensure a level playing field for startups to provide the best capabilities at the best prices to government customers, particularly relevant for programs like CCA.

4. Introduction of  the Adaptive Acquisition Framework (AAF) and the Middle Tier Acquisition (MTA) pathway: the AAF is a new and streamlined approach developed to modernize the acquisition process for defense technologies and systems passed in 2020. It features multiple tailored pathways to accommodate different types of programs, particularly those that need to move faster than the traditional bureaucratic process now known as Major Capability Acquisition, allowing for quicker deployment of critical capabilities. Of particular relevance is the MTA pathway, which allows for Rapid Prototyping and Rapid Fielding of novel technologies, which is more in line with how innovative companies quickly iterate and test prototypes to achieve cutting edge results.

5. STRATFI, OSC, APFIT, and other sources of funding from the DOD to help survive the Valley of Death: In addition to changes like OTA/FFP contracting and MTA acquisition to make it easier to acquire tech from innovative startups and in some ways “shorten” the valley of death, the DOD has stood up multiple programs to ensure these companies receive funding to survive the valley of death. Strategic Funding Increase (STRATFI) awards provide funding to startups after SBIR Phase II awards to facilitate the commercialization of innovative technologies. It requires matching funds from defense or industry partners, which typically aligns itself well with large Series A or Series B raises (you can get a maximum of $60m from a STRATFI award with $30M needing to come from private investors). In addition, the Office of Strategic Capital (OSC) was recently spun up to provide affordable loans to support 14 Critical Technology areas for national defense. Other funding sources like Accelerate the Procurement and Fielding of Innovative Technologies (APFIT) can provide $10-50m to accelerate procurement of systems ready to be fielded. These various sources of funding complement earlier reforms to help startups survive the long procurement times involved in the valley of death.

With these reforms, we have witnessed a shift from a transactional, exclusive process to a more inclusive community of mission-focused players. Organizations like the Defense Innovation Unit (DIU) open channels between startups, investors, end-users, and program managers. The Department of Defense (DoD) has moved from often waiting twenty years for a single solution that fit the traditional mold to engaging with the country's most talented technology players in iterative processes. The established players that once controlled the market are now experiencing shocks of disruption (e.g., Anduril CCA).

It is this preponderance of factors that has recently solidified defense as one of the most exciting areas of venture. From the doldrums to the forefront!

We're fortunate at 8VC to have been involved early with many of the best companies paving the way forward in defense technology. We hope to continue partnering with great founders to employ the most innovative technologies in defending this great nation. As an Easter egg, here's a photo of me with the first-ever Anduril tower!

Interesting things to read/watch

Francois Chollet, Mike Knoop - LLMs won’t lead to AGI
Dario Amodei CEO of Anthropic: Claude, New models, AI safety and Economic impact

ᐧ

In notable news, both InflectionAI and StabilityAI appear to have stumbled, with the former being acquired. While both products may continue to exist in some form, I believe these situations highlight an obvious reality: there is a lack of durability in the value of a Large Language Model (LLM), as they are quickly depreciating assets. In this sense, pure LLM companies seem to be following a path more akin to memory manufacturers: high R&D costs for sustaining investments, low switching costs, and long-term low margins. This is a shift from my previous mental model, where I thought they would more closely resemble real estate investments in growing areas, characterized by high capital costs but long-term steady cash flows and persisting terminal values. (This doesn't mean foundation model companies will be valueless; rather, I think much of their durable value will come from their services, software, infrastructure, and tooling.)

Personally, it's been a delightful quarter in San Francisco. There is an abundance of optimism in the community centered around AI and all the new possibilities emerging. It reminds me greatly of being here during the 2012-2014 period. Far fewer tourists are coming simply because "tech is hot." Instead, young founders and engineers are flocking here because they are passionate about technology and the new opportunities that LLMs are creating. It's a welcome change after the COVID lull to witness the rebirth of tech in the Bay Area. I feel blessed to be a part of it and to experience this pivotal time in human history—in many ways, it must be similar to what our predecessors, like the Traitorous 8, felt during their platform transition.

VC Market Update:

Happy New Year! It's 2021! Well, at least for some of the market. The growth market has been a tale of two markets, with the haves and the have-nots. If you're a fast-growing company in AI or showing signs of category leadership and capital efficiency in another area (such as SaaS or defense), it's like 2021 again. The Information had a good article outlining this, noting that the top 8 LLM-oriented companies have raised at an average of 83 times their Annual Recurring Revenue (ARR) multiple.

This intersects with a lot of what we're seeing -- and in fact, some of what we see in AI is actually crazier than that.

In regular Software-land, we are also seeing some fairly lofty valuations, with growth rounds for efficient and fast-growing companies pricing at 40-80 times their revenue in some cases (I won't name them)! This is happening with multiple term sheets, often up to 6-10. I believe this is a result of the lull we've seen over the last two years and the overabundance of venture capitalists. If you hire investors, they, well, invest. When growth has been scarce for 2 years, there are a lot of people sitting on their hands who will move, and move quickly. I said 2024 was likely to be a great year to fundraise, and I think that holds true.

However, make no mistake, this is not the case for everyone. When I pressure-tested these thoughts with some trusted entrepreneurs, not all of them saw what I was seeing. If you aren't AI-adjacent and don't have particularly fast or efficient growth, it can be rough out there. In normal cases, your valuation multiple has adjusted down to reality, and it can be challenging to raise funds. But at least, in contrast to 2022/2023, there are some significant pockets of life.

What's amusing, though, about the 2021-ification of part of the growth market is that it's significantly more absurd to be doing this today than it was in 2021. At least in 2021, the public market multiples were high; today, they have returned to the old normal. As I mentioned before, when there is an excess of capital, the remedy is that capital must be destroyed!

Macro

To start my Macro update, I will point out that my previous inflation prediction proved accurate more quickly than I anticipated. The market is now pricing in just 1-2 rate cuts this year, with a 20% chance of no cuts at all. This shift in expectations is due to weak GDP data and persistently high inflation. GDP growth for Q1 came in at a disappointing 1.6%, significantly below the expected 2.5%. Meanwhile, inflation remains elevated, with CPI at 3.5% and Core CPI at 3.8%. The Fed's preferred inflation measure, PCE, rose by 3.7%, although I don't put as much stock in this metric.

The key takeaway is that inflation has proven to be quite sticky, and with growth slowing down, we may in the future face the risk of stagflation. Stagflation is a challenging economic condition characterized by low growth and high inflation, which limits the tools available to policy-makers. In this scenario, the only way to combat inflation would be to deliberately trigger a deep recession, a course of action most would prefer to avoid. We're not there yet, but it will be a risk in the future.

Why has inflation proven to be more persistent than most analysts anticipated? As I mentioned in my previous update, several factors contribute to our current inflationary environment: monetary, fiscal, global and even structural (deglobalization). These factors combine to create a "perfect storm" for inflation, making it more challenging to control and likely to persist for longer than initially expected.

Despite stagflation indicators, the markets have shown surprising resilience. The 10-year bond yield has risen significantly, from around 3.9% at the beginning of the year to 4.7%. In spite of that sell-off (bond prices move inverse to the rate), the NASDAQ remains near all-time highs. This is an unusual situation, as growth stocks are generally inversely correlated with interest rates. When rates rise, growth stocks tend to underperform, as their future cash flows are discounted at a higher rate.

I believe the strength of growth stocks can be attributed to an increase in liquidity, driven by the draining of the RRF (Reserve Repo Facility) and TGA (Treasury General Account). As these facilities are drawn down, the excess liquidity is finding its way into the stock market, propping up valuations. I expect this liquidity to persist for most of the year, supporting growth stocks even in the face of rising interest rates.

Bio

The most significant development in the Biotech sector this quarter was the highest level of public market equity issuance since 2021. This indicates that public markets are becoming increasingly receptive to Biotech companies, marking a substantial improvement from the past two years. Although we haven't observed significant changes in the private market yet, this increased public market activity is a promising sign for the future of the Biotech industry.

AI - Robot Foundation Models

This quarter, several AI developments stood out as particularly significant. While there is a plethora of activity in the LLM space, I want to first focus on the innovations happening in the robotics market.

While current language model innovations will undoubtedly be disruptive to the knowledge economy, I believe the greater disruption will occur when robotics revolutionizes labor and the physical world. This quarter, the robotics market made several advances through the funding of Robotic Foundation Models (RFMs) — pretrained, generalized models that can embody physical intelligence and be agnostic to form factors. Although we are still far from seeing these models deployed in the field (perhaps 5 years away from a GPT-like moment), the progress is nonetheless impressive.

To give you a sense of what I mean, take a look at the video from my friend Ilija in the following tweet:

He trained a transformer in (mostly) simulation and deployed it to the real-world zero-shot. Pretty cool!

There is a lot of work in this direction, which I'll summarize below:

• Physical Intelligence raised $70m - Scarce info, but developing foundation models for physically-actuated devices. A video of Chelsea’s work on ALOHA is here.

• Skilld is rumored to raise $300m - Focused on developing RFM for quad and biped robots. The founders seem to be very strong, coming out of CMU. Deepak posted a public video which may show a bit of what they're working on 

• Figure raised $675m - They make humanoid hardware but are rumored to be working on a model too.

• Covariant launched RFM-1 - Covariant released a multi-modal model which can be used to teach and task the model to output action instructions for arbitrary bin-picking.

• Tesla Optimus to Launch by Next Year - Optimus is also both hardware and software. I'm not positive they are training a foundation model, but rumors suggest they are. Video here.

• NVIDIA Announces Project GR00T - NVIDIA jumped in the ring with GR00T, where they will train foundation models for humanoid robots. It seems they will heavily train it on simulation, which could benefit from their Omniverse platform. They also announced a SOC for robotics, Jetson Thor, suggesting they are heavily investing in the physical world.

There's slight variation in approach (humanoid-focused, manipulation-focused), but the ambition is real: models to automate physical work.

These models can potentially have a greater impact beyond just labor. There's an ongoing debate in the AI community on whether LLMs could truly develop higher-order reasoning or AGI (Artificial General Intelligence) based on language alone. The argument is that much of the knowledge embodied in our minds is physical, not just conceptual. As children, the first things we learn are rules about the world around us, not necessarily higher-order concepts or tasks. A thought-provoking podcast by the renowned researcher Jitendra Malik delves into this academic debate with Pieter Abbeel.

In my view, RFMs are going to be big; the question is one of timing—will it take 5 years or 10 years? My bear case argues that these RFMs will, over time, only eliminate the need for most physical labor in society. The bull case, on the other hand, posits that RFMs are the path to achieving AGI and higher-order reasoning.

AI - The real meaning of Devin

For those who missed it, there was a particularly interesting launch this quarter of a company named Cognition. They released an autonomous coding agent that they described as the world's first AI software engineer, called Devin. If you haven't seen it, I think it's worth watching the quick launch video they provided.

The company had one of the most impressive launches I've ever seen for a waitlisted product, making it a household name in the developer community overnight. These types of extreme hype-cycle launches always make me skeptical, and I started off feeling that way about Cognition as well.

However, after trying Devin, I have to say it is incredibly impressive and an important proof of concept for the future of AI-driven application software. I know that's a pretty bold statement to make, especially for a product that most people can't try yet. But I thought I'd share my surface-level observations on what I think Devin does well in advancing the field of AI-enabled application software and how it has illuminated the path forward for software design.

First, I think calling Devin an "AI Software Engineer" is actually a misnomer. I definitely believe this is the ultimate vision, and that it'll get there eventually, but in many ways, the core of Devin's success lies in its design as a system that harnesses man-machine symbiosis. In our tests and by design, Devin is incredibly powerful but does have limitations. We only tried it on net-new software/scripts, whereas most software engineering involves maintenance rather than creating entirely new code. Requirements are input to the system via prompt, whereas in reality, they often encompass product requirements documents (PRDs), other spec documents, and Figma mocks. Lastly, Devin often doesn't get projects perfect; in most tests, Devin wrote a significant amount of code and got things mostly right, but rarely exactly right.

So, what is awesome here?

Even in isolation, while not perfect, you first have to recognize that the capability slope is amazing. In well under a year, the team has created enough tooling to beat other models by at least 3x on the SWE-Bench benchmark. That's pretty incredible. Imagine how good this will get with a few years of continued development.

Next, I believe the design of the system to enable man-machine symbiosis is part of its genius. Devin exposes a very intuitive interface that allows the human engineer to easily see what Devin is doing and correct it in real-time. On the left side is a chat interface where you communicate with Devin, and on the right is your control panel/canvas, which allows you to see what Devin is doing and provide feedback to steer him down the right path when he does something wrong or gets stuck. The UI serves as your command center. Devin is your pair programmer rather than a standalone engineer. The human acts as the grader, while Devin acts as the writer. This paradigm is likely to be adopted by most highly-performance AI apps in the near term, where the models we use are mostly right but not exactly right.

Lastly, I believe Devin demonstrates what the future of interesting "thick-layer" LLM-enabled applications will look like. To make the system work, Cognition had to build a lot of supportive multi-agent tooling and provide Devin with enough extensions to perform proper planning, reasoning, and possess the necessary skills to interact with the outside world (e.g., posting on developer forums, searching codebases for context, etc...). This contradicts my previous update, where I thought integrating LLMs into applications would be straightforward, requiring only smart prompting. In reality, harnessing the true power of LLMs might require building fairly complex tooling around the system. This insight is quite profound, as it opens up the door to much more potential for a) disruption of incumbent systems that fail to get this tooling right, and b) an archetype of AI company where there is a great deal of complexity/defensibility that "looks" a lot like a typical SaaS investment.

In one fell swoop, Devin has rejuvenated the multi-agent space, which until now was quite unpopular. I believe there's a real opportunity for people to create these types of helper-AI systems in every domain. Wherever there is knowledge work, I think there will be grader-writer systems that enable man-machine symbiosis to give humans superpowers. In most domains, it will be a race to see if incumbents can build these systems before new entrants bring them to market.

AI - Model convergence

This quarter was a very important quarter for language models. I'd argue it was the most significant quarter after the ChatGPT release. We saw notable releases from Databricks' DBRX, MistralAI's Mistral 8x22B, Cohere's Command R/R+, Llama 3 from Meta, and interesting models from RekaAI.

Most notably, I believe the perceived infallibility of OpenAI was challenged, as for part of the quarter, Anthropic's Claude 3 Opus outperformed GPT-4-turbo, perhaps marking the first time OpenAI was not the clear leader (this has since been resolved, as it seems GPT-4's performance had temporarily degraded mid-quarter). I think we'll see a new release from OpenAI soon (possibly GPT-4.5) that provides a significant performance boost, but it's certainly true that their competitive edge has eroded.

While it's arguable who was the best model now, I think it's not arguable to observe that everyone is getting *really* close to one another, including Llama 3-70b which is actually Open-Source. This widespread convergence of models has happened much much faster than I anticipated and I think really brings into question the durability of value captured by the LLM-layer itself. Especially since switching costs are very low.

There are also questions posed on the nature of competition and who will ultimately lead the LLM race. The traditional view is that we are in a world where all that matters is data and compute; those with the most capital (the hyperscalers or those who can raise seemingly infinite amounts) are likely to meet at the top. Reka, model #13, may be an interesting counterexample. They have publicly only raised $50m but are competing with much more resourced competitors.

I have a sneaking suspicion that as we move forward over the coming decade, more novel mid-quality data and increased compute for training will not drive the major gains in creating the best LLMs. Some scale will be table-stakes especially with current architectures; but aside from models having logarithmic scaling, I believe this because we are already discovering more compute-efficient ways to train models, and this trend will likely continue. This makes intuitive sense to me; if we consider how humans learn, we don't need to read the entire internet to acquire knowledge. Similarly, on the data side, we probably already have all the data necessary to learn most human knowledge—the frontier likely lies in either making better use of existing data or creating synthetic data from high-quality sources to enhance learning. I bet the future step-function advances will be algorithmic -- novel architectures, tooling and compute optimization will reign supreme and eventually increase competition further.

Now, it's certainly possible that proprietary data sources will help drive some differentiation. I can see the argument for a network effect driving better fine-tuning via Direct Preference Optimization as a potential layer of proprietary performance enhancement. Or high-quality proprietary data sources in general to make custom-models.

But I think on balance given the speed of model capability convergence, and the ease of switching, it appears that eventually the LLM-layer will be a commodity product. True convergence probably won't be seen in the short-term while there will be lots of iteration on architecture and tooling to drive outperformance, but as all of that stabilizes, LLM-capability won't drive much profit. That is, unless we make significant advances on the tooling or architecture layer that are not broadly disseminated by a current player (OpenAI?).

Despite these competitive considerations, the LLM funding market continues to thrive, with Open Source Mistral apparently raising funds at a $5 billion valuation even after the release of Llama 3! I don't quite get the mania -- especially since Meta has made it clear they're willing to play in Open Source for a long time.

Conclusion

I conclude this update feeling both optimistic and humbled. The technological frontier is evolving at a rapid pace, and many of the beliefs I held in the previous update are no longer true. This is partially due to unpredictable changes and partially a result of further reflection on my part. In either case, I'm going to strive to be honest and unattached to prior views when updating my perspective. To be a great innovator, I don't believe you always need to be right, but you must recognize when you're wrong and adapt your viewpoint more quickly than your competitors.

PS - For those waiting to read my thoughts on natural growth rates, I've decided to upgrade (or downgrade?) that topic to a standalone article.

Interesting things to read/watch

Scalable AI Architectures by Kevin Niechen @ 8vc

Whitney Baker on the Death of American Exceptionalism

Jeff Dean (Google): Exciting Trends in Machine Learning

Jitendra Malik: Building AI from the ground-up, sensorimotor before language

Fortunately for most mega-funds, the extreme capital inefficiency of foundational AI infrastructure companies aligns well with their business need to deploy mega-funds biennially. It's a true match made in heaven! We'll see if such a strategy yields expected returns. I remain a bit skeptical and generally bullish on model bifurcation: OpenAI for high-complexity use-cases, Open Source for mid-range applications, and little in between.

Venture Staffing 

To me, the real story of this quarter is not so much about the market for deals in Venture Capital (VC), but rather about the market for people. My view is that we are likely to witness a significant exodus and shrinking of VC firms over the next 2-3 years. I've expressed this perspective in a well-timed tweet recently:

My tweet preceded/was coincident with a few big departures that got announced:

• The head of Gradient ventures leaving

• The entire fund OpenView shutting down!

• MG Siegler leaving Google Ventures

• Mark Goldberg leaves Index Ventures

And there are several other unannounced departures I won't disclose here, but I can share if you're curious. I believe this is a trend we'll see over the next few years for a few reasons:

1. Performance: As we emerge from the shadow of 2021, we are likely to witness significant portfolio losses, potentially leading to the end of many careers. A mentor of mine once said, 'The remedy to an excess capital environment is that capital gets destroyed' (h/t Patrick Kenary). Consequently, following the excesses of 2021, numerous careers may also come to an end.

2. Shrinking Dealflow/Market: We are witnessing the culmination of several technology waves: mobile, social, and cloud. These sectors have experienced a secular decline over many years. This trend seemed to reverse temporarily with Tiger's influx of liquidity in 2021, but this only created a false sense of vitality in the innovation economy. While BioTech hasn’t been declining secularly, it has certainly peaked after the 2020 sugar-rush and is now undergoing significant retrenchment in company formation as well. Currently, Artificial Intelligence is emerging as the new trend, rejuvenating the market. Despite the AI renaissance, I estimate that the opportunity for venture investment is still about 60-70% smaller than during the 2018-2021 period, which implies a reduced need for VCs.

3. Beginning of Cycle dynamics: The 10-year technology bull market has ended, reaching its zenith in 2021. Many VCs made significant profits and, frankly, lack the will or desire to engage in another decade-long cycle at the start of the new bull market. The AI wave will build a new generation of fortunes, but as always, it will be a messy, hard, competitive process, requiring 'hand-to-hand combat' with limited immediate rewards. For some, the drive to continue will wane – after all, life is finite.

Personally, I'm in the grindset. I'm fully committed and eager to tackle the challenges of transitioning from one supercycle to the next, embracing both the challenging ending and optimistic beginning it brings.

According to PitchBook data, the trend of reduced VC staffing appears to be underway. I wouldn't be surprised if, within the next 1-2 years, we see numbers falling below those of 2018

The number of investors engaged in deal-making has decreased by 38%, a trend distinct from the 72% reduction in funds raised this year. This fund decline is more a result of lower deployment rates rather than a decrease in the number of investors.

Fundraising Market

Overall, this quarter, public market software multiples have slightly increased, based on the 8VC Public Software Index. This mirrors what we've seen in the private early-stage market since last quarter as well. The growth market has continued to be subdued, not helped by the mixed IPO performance of the three major Q4 IPOs.

By the numbers, 2023 in fact has remained a fairly quiet IPO year, well below-trend:

But I believe that this scenario is set to change. In 2024, I anticipate that more companies will recognize public markets as a credible path for raising capital. This could circumvent the challenges associated with structured financings or down rounds in the private markets. An IPO leads to an automatic conversion to common stock, which clears out the preference stack and simplifies a company's financial structure. Additionally, it necessitates improved governance, which could be leveraged as a quid pro quo for approving a down-round IPO.

While there is a prevailing 'wisdom' among bankers and pundits that 'the IPO window is closed,' I have always considered this perspective to be rather simplistic. Various excuses are often cited, such as 'not enough analysts,' 'banks won't build the book,' or 'the last few IPOs didn't pop.' However, in my view, these are merely superficial reasons used by the uninformed to appear knowledgeable. The reality is that price clears all markets [and preference stacks]. What constitutes a good IPO at one price can be unfavorable at another. The financial system is flush with liquidity, and where there is capital, there is opportunity. I believe that more companies will follow Instacart's lead in going public at lower prices, making the issue an appealing buy.

In the early-stage market, a particularly notable observation I have made is that the majority of our investments in 2023 have been predominantly in Seed rounds. This scarcity of Series A funding is a trend also observed by many of my peers. Unsurprisingly, the data shows a noticeable and industry-wide decline in the conversion rate from Seed to Series A investments since 2022.

Aggregate statistics bear out that Seed is the most resilient part of the market today.

In my opinion, the sobering reality behind the softness of the A+ VC market is straightforward: companies simply aren't performing well. As my tongue-in-cheek subtitle suggests, growth has been slow across many companies in the private markets, though there are certainly outliers. For example, I sit on the board of a company that is growing at 250% this year while generating nine figures in revenue. However, for most companies, the market for selling software has been sluggish throughout the year. The graph below illustrates growth challenges in the public markets.

The numbers provide some relief in Q3, but over the past year, growth in the software sector has faced challenges. The above chart, which measures the second derivative of public company ARR growth—specifically, Net ARR Added YoY—should always be positive if companies are growing exponentially. Anecdotally, in the private markets, I've observed that some companies are not experiencing any growth at all. Several factors could be contributing to this: strained corporate IT budgets, the rise of new tech incumbents with strong technology cultures, and the intense competition in the software market. Next month, Joe and I plan to explore this topic in more detail. Perhaps my most controversial viewpoint is that COVID-19, which spurred a 'great resignation' among high-performance collaborative cultures and a widespread shift to remote work, has slowed down nearly every company.

The unfortunate conclusion is that, with growth under pressure, I believe 2024 will surpass 2023 in terms of startup shutdowns, particularly among later-stage companies. We witnessed this trend in 2023, and it is likely to worsen in 2024:

The downside of the Venture Funding Markets is that many companies are struggling. However, the upside lies in the abundance of dry powder and the presence of numerous junior partners who are eager and ready to invest in great opportunities. The growth challenges we observe are not due to a lack of capital, as there is plenty available. If you identify a unique market opportunity, demonstrate strong capital efficiency metrics, and achieve solid benchmarks, you are in a good position to raise a substantial Series A round, which is likely to be competitive. We still see top performers out there. Companies not operating in saturated markets are performing well. If you aren’t in one of those markets, the key is to focus on managing burn rate. In my next update, I’ll discuss more about managing your company’s natural growth rate.

Lulls in growth are temporary and often pave the way for cyclical upswings. Don't lose hope or become overly focused on market updates — instead, strive to become an outlier!

Macro

The big news to end the quarter in the macroeconomic landscape appears to be 'The Fed Pivot' — indicating that the Federal Reserve has subtly signaled its intention to start cutting interest rates. In response, the market has obliged by pricing in six rate cuts for next year:

On the surface this makes some sense because we're seeing inflation ease, and based on money growth this should continue into 2024:

Up to Q3 2023, the money supply has been constricted through various means: Quantitative Tightening (QT—where the Fed shrinks the balance sheet), the Reverse Repo Facility (RRF, the Fed removing reserves from the banking system via repurchase agreements), and the build-up of the Treasury General Account (TGA, which also removes reserves from the banking system). These actions have contributed to a decline in M2.  In Q4, with the exception of QT, these trends have started to reverse, primarily due to extreme deficits effectively injecting more liquidity into the system (drawing the RRF and TGA). The TLDR; is, much of what has led to M2 declines is starting to reverse, suggesting inflation pressure in 12+ months.

Beyond just monetary factors driving lower inflation, another element contributing to the decline is the recession occurring in the developed world outside of the US. This downturn is helping to suppress goods inflation, which has been a major factor in the overall reduction of inflation. However, this situation also poses a risk: cyclically, as the rest of the world begins to recover and grow in 2024, it could lead to renewed inflationary pressures. This is particularly concerning if it coincides with the Federal Reserve starting to ease its policies.

These coming inflation pressures (DM growth, RRP/TPA declining, strong US growth) coincide with massive, continued stimulatory government deficits, as illustrated below.

In summary, I believe that due to a confluence of factors, it is inevitable that inflation will resurface in late 2024 or early 2025. This is particularly likely given that core inflation currently stands at 3.9%, significantly higher than the target of 2%. The Federal Reserve's eventual decision to cut rates will likely be a mistake. Why would the Fed make such a mistake? It's primarily because inflation is expected to trend lower early in the year, coupled with a strong incentive to stimulate the economy ahead of the election.

In all likelihood, the Fed's easing in 2024 will present a favorable opportunity for founders to raise capital, as SaaS multiples will expand and liquidity will increase. However, I believe this situation will be short-lived. Therefore, 2024 is likely to offer a valuable, albeit brief, window for fundraising.

BioTech

In this quarter, we observed notable activity in the public BioTech markets, particularly evident in the rally of the XBI. This rally was largely driven by M&A activity, with major deals such as the acquisitions of Karuna and Carmot standing out, valued at $14 billion and $2.7 billion, respectively. Year-over-year, M&A activity surged by 79%, totaling $152 billion, with $63 billion of that occurring in Q4 alone. This upward trend in M&A is likely to persist as Pharma companies aim to replenish their dwindling drug assets. An important factor in this trend is the $236 billion in Pharma revenue expected to come off patent between now and 2030, presenting a substantial liquidity tailwind for successful BioTech companies. On a personal note, I bought some XBI at the end of October — I’ll make sure to Tweet such investments in the future!

Even amidst the market rally and generally positive sentiment around M&A and partnerships, the BioTech sector remains subdued. As of EOY 2023, 169 public BioTech companies are trading at a negative Enterprise Value—they are worth less than cash. Put another way, many BioTech companies are so poorly valued that you can't give them away.

Sadly, I believe the private BioTech markets are in for a rough ride unless there's a significant increase in valuations, which, realistically, I don't anticipate happening. However, a challenging market does not necessarily mean a lack of investment opportunities. Often, it is during these rough times that the best investment opportunities arise — but realizing returns in such a market will be challenging.

Why do I believe there won't be a significant increase in valuations? As outlined in my Q3 note, I think we have entered what can be described as a new normal in valuation regimes, closely aligning with the environment before 2020. Notably, the XBI, although not my index of choice, is trading around the pre-2020 mean. This level, in my opinion, accurately reflects the current state of private market valuations as well. From my perspective, the valuations today, last seen in 2018, offer a more intrinsic representation of true therapeutic expected values, especially when compared to the inflated figures observed in 2020 and 2021.

Coming out of JPM, I’ve encountered a mix of positive and negative sentiments. On the positive side, Pharma companies continue to demonstrate a strong appetite for deals, either through partnerships or M&A. On the negative side, there are notable struggles in the private BioTech markets, with many companies quietly shutting down. This shutdown trend seems intuitive to me, and I anticipate witnessing even more shutdowns, regardless of whether valuations broadly increase.

Why do I think the private BioTech markets face major challenges? There are two primary reasons. Firstly, the excesses in the public BioTech markets during 2020 and 2021 led to a surge in the founding of weaker companies, leaving many private companies stranded. BioTech investing during this period resembled financial engineering more than traditional VC. The typical company-building strategy involved spinning out some interesting preclinical IP from a lab, assembling a credible management team, and then debuting it into the public markets at the preclinical stage — distribute to your investors, rinse, and repeat. I have always been skeptical of this model as an effective recipe for value creation. In my opinion, investors should commit to funding their companies through to clinical readouts while remaining private. Additionally, there had been considerable mimicry in target selection, which is likely to result in many inferior “me-too” assets of limited value. For example, ClinicalTrials.gov lists over 188 trials actively recruiting CD19-positive patients. Ironically, it’s the most crowded areas that are the riskiest and least value-accretive.

Secondly, many private BioTech companies are grappling with cost structures that don’t align with current market valuations. At our fund, I have been a vocal advocate for increasing our investment in BioTech, particularly in light of the recent decline in prices. However, we often encounter a significant challenge when considering investments at a company's proposed new post-money valuation (pre-money valuation + money in the round). Often, when comparing to similar companies, we find that the next funding round for these companies is likely to be flat or even down, despite them achieving significant milestones. Their burn rates are often expected to exceed the value they create.

Fundamentally, private BioTech companies need to improve their operational efficiency. It's common to see these companies recruiting numerous executives from Big Pharma, who often tend to establish a Pharma-like structure internally. This approach can lead to layered management and excessive staffing. In a startup environment, it is crucial for every team member to actively contribute. Additionally, significant advances in infrastructure have made therapeutic discovery more cost-effective. Companies should consider leveraging outsourced, cost-efficient platform discovery services, such as AlloyTX, to boost efficiency and reduce headcount. Such services can provide essential support while reducing overall burn.

Achieving high efficiency in the BioTech sector is indeed possible, especially in founder-driven companies. For instance, I have witnessed a BioTech company successfully advance a cell therapy to trials with just $40 million in funding. In another case, an antibody platform company efficiently delivered a Phase 1 asset with clinical readouts, supported its platform with only one partnership, and developed another IND-ready drug, all within a total budget of $60 million. Impressively, both companies have reported very positive outcomes from patient trials, suggesting capital expense is not always correlated to clinical efficacy.

Unfortunately, as the excess in the private BioTech markets is winding down, fundraising has become increasingly challenging. This is particularly true for early-stage funding, where most investors are now focused on supporting their existing portfolio companies through flat or bridge funding rounds, rather than exploring new opportunities. For example, we are closely involved with two companies that have been trying for 12 months to secure their Series A funding round, illustrating the heightened difficulties in the current investment climate.

The fundraising situation in the BioTech sector is unlikely to see immediate improvement, particularly in light of the potential shutdown of BioTech VC firms who were active in the recent 2020 investment frenzy. This could further deplete an already limited capital pool. Despite these challenges, I am hopeful for two positive developments: Firstly, a broader realization within the BioTech market that capital is not unlimited, which could prompt a reset in cost structures. Secondly, the closure of unviable companies may allow valuable talent to reenter and enrich the ecosystem.

In the midst of the prevailing doom and gloom, I remain steadfastly optimistic about lean BioTech companies that are adept at navigating the current environment. A significant benefit of the ZIRP era was the discovery of numerous ‘0-1’ innovations, spurred by the funding available during that period. Once translated, these innovations are likely to produce many groundbreaking clinical drugs. The next generation of BioTech companies, built upon this wave of innovation and focusing on cost-effective scaling, are well-positioned to capitalize on this investment. Moreover, the impending patent cliff in Big Pharma — with an estimated $236 billion in revenue expected to come off patent by 2030 — presents a substantial opportunity for companies that can demonstrate clinical proof of concept. Entrepreneurs and investors who are efficient in developing and delivering effective drugs are likely to find an abundance of capital for exits.

AI

Wow, this note is much longer than I had hoped, and we're just starting to discuss AI. I have a feeling that AI will be a frequent topic in these writings. The first thing I'll cover is what type of disruptive technology I believe AI represents.

Let me set the expectations upfront: I'm e/acc and I'm a believer. I wasn’t among the earliest to recognize the potential of Transformer Architectures, but using ChatGPT in December 2022 was a groundbreaking experience for me—the first time a new technology truly awed me. Having studied AI/ML extensively in the 2000s and early 2010s, I initially believed it would be valuable but not necessarily revolutionary. Now, I am convinced that Large Language Models (LLMs) will drive fundamental change in our society. In my opinion, we are on the cusp of the next industrial revolution, driven by AI. I anticipate that AI will be over-hyped in the short term (1-3 years), a period likely marked by struggles in AI infrastructure companies and disappointments in AI deployment, which I call the '90/10 problem.' But looking ahead, I believe that in the next decade, AI will dramatically transform a significant portion of knowledge work.

What changes do I foresee with the advancement of AI? Here are some key shifts that come to mind:

• The landscape of commercial go-to-market strategies will undergo a radical transformation. As a buyer you'll get an overwhelming number of perfectly personalized AI-generated emails and phone calls. As a result, buyers will likely start to rely more on referrals, in-person tradeshows, and trade associations for trustworthy product recommendation

• Professions that primarily involve synthesizing mostly structured knowledge are set to diminish. This includes careers like programming, law, consulting, and certain medical practices, which will gradually become less prevalent.

• A key area of human defensibility will be in aiding human interactions with AI in complex or emotionally charged situations. For example, the lawyers who adapt will focus on gathering context to feed into AI lawyers. Similarly, Doctors will work with patients to gather context and explain complex or emotional realities

• So far, LLMs have primarily targeted digital and knowledge-based domains. Disruption in physical labor by AI, with pre trained Robotic Transformer Models, is likely further down the line (I am currently exploring this area and would welcome introductions to anyone working on it).

• Ironically, those who told coal miners to learn to code, will be the ones out of a job before the miners themselves; both journalists and programmers!

I often wonder what the future will look like in 50 years, once LLMs have fully integrated into our lives and automated knowledge-work. The somewhat ironic Marxist perspective suggests that we may shift towards more community-oriented living, as most of the remaining work will be on a human scale. If you visit a village in southern Italy, you'll notice that most activities there would remain largely unchanged in a post-LLM world. I believe our future might more closely resemble this kind of lifestyle.

Now, let's address the questions: Are LLMs a platform? And do they represent a sustaining or disruptive innovation?

The first answer depends on your definition of a software platform. If we consider whether new software systems will be built on top of LLMs as their underlying hosted platform of record, technical framework, or UI, then the answer is probably no. For the foreseeable future, I believe LLMs will primarily serve as additive features/services to software systems built on existing platforms. Even once Artificial General Intelligence (AGI) is achieved, the AGI system itself will likely become the ultimate core system of record. This is because it would be all-knowing and capable of building any supplemental infrastructure it may need, rather than serving as the foundational layer upon which a software system is structured.

I believe that LLMs offer a significant advantage to software vendors who already have an established customer base. Consider a scenario where you operate a title insurance franchise and wish to automate much of your document processing using LLMs. In this case, it's more likely that you would opt for AI features from your existing System of Record (SOR) vendor, assuming they have integrated such capabilities, rather than acquiring a completely new system built from scratch to incorporate LLMs. The primary reason for this belief is the inherent ease of integrating LLMs. By inputting the right context and the appropriate prompt in English, a developer can tap into most of the capabilities that the LLM offers for a specific workflow. This suggests that even legacy players in the industry should find success in integrating these technologies.

As to whether LLMs are disruptive or sustaining innovations, I believe the answer depends on the market in question, given their broad applicability.

For enterprise Independent Software Vendors (ISVs), as discussed LLMs are currently a sustaining innovation.

For some consumer internet companies, like Google, whose products rely on interpreting, synthesizing, or reasoning based on public knowledge, they are disruptive innovations. LLMs are a step-function improvement but at a much higher (real) price-point. (Perplexity.ai is a way better product than Google but more costly to run a query). If costs come down, then existing consumer franchises can be threatened (i.e. where there is less of an install-base lock-in and more a brand-effect).

For white-collar workers, AI represents a disruptive technology. At present, LLMs primarily compete in tasks of lower quality and with limited reliability. However, over time, they are likely to disrupt many aspects of the work performed by knowledge workers, as previously described.

In the very long term, there's a possibility that LLMs could disrupt enterprise ISVs. If LLMs evolve to translate requirements into fully-fledged systems (including migration scripts), then every software sale might face competition from a (less expensive) custom build via an LLM. This scenario would massively disrupt enterprise software pricing. However, my reservation lies in three counter-arguments: a) as Steve Jobs pointed out, there's always room for taste in products, b) ISVs offer more services than just software, and c) most customers don't quite know how to describe what they want. If we harken back to the old-days of Software Engineering in the 80s/90s, half of all internal software projects would fail. A large part of that was due to requirements gathering and design being a real challenge -- not implementation!

So, there's a lot of theory, but where are the opportunities? I believe that the first-order effects of AI are already priced in—when looking at foundation models, consumer products, or certain types of AI infrastructure, there is no alpha. [As an aside, AI is probably one of the few areas where alpha was priced out even when it was somewhat contrarian, as many hot companies had seed rounds at valuations of $100m+ pre-money].

In my opinion, the second-order effects are where the opportunities lie. In no particular order and non-exhaustively:

• AI-enabled services - This is probably the area I'm most bullish on at the moment. Consider existing businesses that provide a service; augment labor as much as possible with AI, and quickly dominate the space with that cost advantage. Key areas are ones which include structured information extraction, summarization, knowledge retrieval, and basic document synthesis tasks. 

• Net-new application layer products (voice, enterprise search, etc...) - As mentioned above, I'm not a proponent of rewriting the existing application layer. However, new application spaces are now feasible with AI. The bar is high; one must ask, 'Does AI fundamentally enable this new product?' Interesting areas include those with unstructured input data, different input modalities (voice, music, images), or where there is inherent bundling of multiple products because LLMs make it easier to integrate workflows together (from prospecting to sales outreach).

• Serverless training/burst GPU computing - Anecdotally, we hear of very low utilization rates for prepaid GPU resources, as low as 10% in major data centers. This inefficiency seems ripe for a solution.

• Open Source - I am a strong advocate of Open Source models (more on this in the next update). Consequently, tooling for deploying, tuning, etc., to make these models more accessible will be highly valuable.

• LLM Evaluation/monitoring/production readiness - Suppose I've developed an LLM application and want to deploy it. Key considerations include: a) evaluating the boundary conditions my model will work within, b) understanding the error rate, c) testing and d) determining if a new foundational model is disrupting my application (i.e., how to test subjective outputs).

• Enterprise LLM application development tooling - Similar to the above, the tooling for creating LLM applications needs enhancement. Enterprises need comprehensive solutions for creating a RAG-stack that respects permissions, chunks data properly, optimizes embeddings, distills a model, and even provides guidance on how to effectively prompt a model, especially one that has been fine-tuned, to achieve the best results.

• Code generation: Coding is likely to be one of the first areas significantly enhanced by AI. There will be room for a variety of tools aimed at boosting internal developer productivity, especially in environments where code security is of paramount importance.

Conclusion to 2024

This update ended up being much longer than I had thought. Whenever I look forward to 2024, I find myself thinking things will be extremely turbulent but optimistic. The natural conclusion of poor startup growth is that 2024 will see more startups go bust than ever (inVision which raised $350m at $2b valuation just died). At the same time, an easing Fed will likely present an opportunity for growing startups to raise money on favorable terms (in public or private market). There will be haves and have nots.

With the upcoming turmoil facing under-performing businesses, I am optimistic about witnessing a cultural and cathartic rebirth that Silicon Valley urgently needs. In my view, COVID-19 brought about a 'Great Resignation' in the commitment of employees to high-performance, collaborative company cultures. This change saw many startup employees lose their deep connection with their work. The Great Resignation has been one of the most destructive yet least discussed trends in the startup economy. I anticipate that, with companies under stress, there will be an end to this period of decadence. And that insane, collaborative, optimistic spark which makes Silicon Valley tick will come back again. 

Interesting things to read/watch:

Founding of Transdigm: 

Karpathy intro to LLMs: 

History of Nvidia on the Acquired Podcast: 

Fundraising market -- return to the old normal

In software, we're starting to see more activity in the early-stage, but late-stage remains quiet outside of a few exceptional companies. Seed has been quite active this quarter with YC in August and my impression is a lot of those companies got a great reception. I've started to hear about more Series As getting done (at 2018 prices), although we haven't been too active ourselves (not for lack of interest though, we are open for business). The late-stage remains a desert. From anecdotes, my good friend at a Top-Tier Growth Fund says they've only done 2 deals in the last 27 months. I don't think this is for a lack of capital in the market, but instead a mismatch of valuation expectations. Instacart seems poised to price ~76% down from their last mark (back to the 2018 price) in the IPO and I think a lot of late-stage companies will be in this spot. It's a lot easier to take this haircut in the public markets where the pref gets wiped, vs. doing so in the private markets. I think this pref dynamic will drive more IPOs next year. 

In Bio, we see the opposite. Early-stage funding is relatively quiet/challenging, and the late-stage market is more active -- with the caveat that it remains very difficult to get external capital unless you are clinical, or funding through clinical readouts. The "next round" in Bio is not a given, so investors want to see their round of funding get through unquestionable value-drivers (i.e. human readouts where possible).

Broadly data seems to show we may be near a bottom in capital deployment:

Source: SVB State of the Markets H2 2023

The dominant narrative I see around SV is that valuations/the venture markets are "depressed" and companies/commentators are encouraging companies to survive until conditions improve. I think this type of narrative -- implying the current environment is an aberration -- belies the fact that the present environment is much more normal than was 2021. In my view, the present environment is the new normal, and the new normal looks a lot like the old normal of 2018.

Why do I think this? Because in many ways public market valuations/expectations are optimistic/aggressive, and public software multiples have rallied back to more normal levels.

If we dissect the data, public software multiples have rallied back from the bottom, back to the mid-2010s average. It looks much more like '14-’18 which is in reality a pretty average environment for Software Multiples.

Source: Capital IQ, Company Filings, 8VC Estimates

Commentary: Public market SaaS multiples are down 33% from the 5-year average, but near the 5-year average of ‘14'-'18

This is all in the face of positive economic data and aggressive earnings (forward PE ratios near 20x -- very high). Will there be a recession? Basically no, at least when you look at corporate earnings. The expectation is that earnings have bottomed in Q3 and will start growing again. No recession here.

Source: IBES Refinitiv, Company Reports, S&P Dow Jones

Commentary: S&P 500 earnings are expected to increase over the next 5 quarters. The market is pricing in fairly optimistic growth — not a recession or slowdown!

Examining the broader market, valuations indicate that the S&P is priced quite aggressively. Instead of using P/E ratios, I prefer to consider the Equity Risk Premium (defined as the yield of equities minus the risk-free rate). This metric excels over P/E ratios in normalizing valuations in the context of interest rates. By this standard, the broader market is as expensive as it was prior to the Global Financial Crisis (GFC). This certainly doesn't suggest things are depressed!

Source: S&P Down Jones, Company Reports, FRED

Commentary: The Equity Risk Premium (ERP) represents the amount of risk priced into equities. A low ERP indicates that stocks are expensive relative to the risk-free yield, typically measured using the 10-year rate. Based on this metric, stocks are currently more expensive than they have been at any point in the past two decades.

What's all this mean? It means a lot of positive news has been factored into financial markets, and the present venture environment stands in the presence of an aggressive market backdrop. We shouldn't expect significant improvements in the venture funding markets from their current state. Everyone should plan accordingly; capital is expensive and your growth needs to be efficient. This was the way of the past, and it's likely to remain so in the future. The old normal is maybe not as fun but it is the way.

What does the old normal look like? If you're a post-seed software company, assume you're going to get 10x 6- to 12- months out ARR if your growth is efficient (good growth rate, burn ratio and magic number). If you're a bio company, you won't be able to fund long platform builds anymore. Your first round needs to get you to an IND, and your subsequent rounds need to get through clinical readouts.

Inflation

A counterargument to my perspective on the 'old normal' is that a Fed Pivot is imminent, which would result in lower rates and another liquidity-driven surge. I think that's very unlikely anytime soon for a few reasons. First, by the numbers, Core CPI was 4.7% in July. This is very high and well above target. We are not close to normalizing inflation.

Second, the Fed is cognizant of the fact that inflation can be very volatile, and unlike the '70s they will be loath to lower rates or step off the gas unless inflation materially abates and unemployment increases (aside: if they forget this fact, inflation will probably come back and force them to raise again anyway). Compare inflation from the '70s to today to appreciate the volatility of high-inflation environments.

Source: Bloomberg

Commentary: The graphic juxtaposes inflation from the past decade with inflation from 1966 to 1982. Although the exact scales might not be directly comparable, the aim is to illustrate that during periods of high inflation, inflation can be extremely volatile and challenging to control. The Fed is cognizant of this history and is likely to maintain higher interest rates to avoid a recurrence of the inflation spike seen from 1978 to 1982. It’s unlikely we will see another liquidity boom like we did in 2020/2021.

Third, most macro drivers point to high inflation. Unemployment remains very low. Most debt in the US is fixed rate (this differs from other countries) and so the effect of interest rate increases has not been felt outside of core cyclicals. The fiscal situation in the US is abysmal, we are running war-time deficits which are stimulating the economy while the Fed is trying to slow the economy. And to make matters worse, the government is one of the only entities that issues a majority of short-term debt, so it is feeling interest rate hikes in the deficit. Lastly, I think the conflict in Ukraine is likely to make energy and soft-commodity markets much more volatile this winter. All of this is inflationary!

AI

The lone bubble spot we've seen in the market has been Generative AI. I'll delve into my AI thesis in the next update, but I genuinely believe that the impact of this new generation of models will be revolutionary over the next decade. It will transform how every industry operates. We continue to see many companies raising incredible amounts of capital at irrational valuations with relative ease, particularly for foundation models and AI infrastructure. 

I believe the majority of the capital invested today will be lost, similar to what happened in the .com era. However, I expect there will be a select group of companies that endure and build the future. One theory I hold is that market cycles now move faster than in the past due to the rapid dissemination of information. I predict the AI bubble will burst sooner than the .com bubble did (1995-2000) -- to be succeeded by a long bull market for Generative AI. It might not happen immediately, but we are already witnessing a gradual waning of interest in Generative AI from peak. Data from SimilarWeb on GenAI traffic shows there are declines across the board (and it’s widely reported ChatGPT revenue slowed in the summer) . I believe that a drop in interest would be beneficial for the sector. [I can’t share this traffic data publicly since it is copyrighted]

Our present interest is heavily concentrated in AI-enabled services. I think the next wave of Vertical Software is AI-enabled services automation. If you know of any labor/services-heavy parts of your vertical market, please send me a note about it.

Conclusion

My note is not meant to be pessimistic but sobering. There is no depression in capital markets, but rather a return to the old normal. Great businesses can be built in any market environment, and in particular, exceptional businesses thrive in loose labor markets. In this note, I haven't made many forward-looking statements, I'll save forward-looking predictions for the next note. But I tend to think most risks in the market remain to the downside. Conditions are likely to get worse than better since we're priced to perfection. It's important, however, to know that great businesses will always endure and can raise capital in any environment -- the correct takeaway is to adjust expectations.

What makes me bullish? 

Bad companies are being washed out and good companies are finding more talent available. A company I'm on the board of had 200 applications for a Sr Engineering role in a week! This rings true for our bio companies as well.

What makes me nervous?

Most companies are still missing projections, and it seems *hard* to sell software. This makes me think the real economy is weaker than employment and S&P multiples imply. 

Until next time. 

Good things I've read/listened to this quarter: