TSMC’s Wafer Scarcity May Be Preventing an AI Overbuild

Demand is arriving faster than the physical, financial, and strategic system can satisfy it. That was the governing frame Gavin Baker applied to the current AI cycle, with Anthropic as the reference point. The company, he said, added $11 billion of ARR in one month.

The comparison was meant to be jarring. Palantir, Snowflake, and Databricks, in Baker’s telling, are three of the highest-profile SaaS companies founded over roughly the past 10 to 12 years. They employ thousands of people and spent a decade building. Anthropic, he said, added their combined businesses in a month.

That claim shaped his interpretation of the market weakness in March and April. Baker distinguished between drawdowns that prove an investor wrong and drawdowns where price action moves against companies the investor believes are being mispriced. March was the second kind for him: the Nasdaq sold off while AI fundamentals, in his view, were accelerating at a pace without precedent.

DeepSeek had created a similar setup earlier in the year, but required more work to read. The initial market reaction treated DeepSeek as negative for the AI trade. Baker read it the opposite way. By “DeepSeek Monday,” he said, it was already clear to him that the release was positive for compute demand: prices in AWS availability zones in Asia had doubled, GPU availability was falling, and GPU rental prices were starting to rise. The lesson was that reasoning models were far more compute-hungry at inference time than non-reasoning models.

Anthropic’s growth required less interpretation. “All you had to do,” in Baker’s words, was observe what was happening. That made the April opportunity more striking to him: investors who regretted not buying in 2022, during Covid, or after DeepSeek had another chance, with what he regarded as a clearer AI inflection and unusually attractive relative valuations.

Patrick O'Shaughnessy pressed on why Anthropic and OpenAI, the purest reference assets for the model layer, did not look more expensive on revenue multiples relative to earlier software peaks. Baker separated the two companies. Anthropic and OpenAI, he said, are “pretty different animals” in capital efficiency. Anthropic has a dramatically lower cost per token, visible in how much less capital it has burned to reach roughly similar revenue scale. He estimated Anthropic may have burned 80% less than OpenAI.

OpenAI, by contrast, has secured more compute, and Baker said that aggressiveness “really paid.” But he still argued that Anthropic’s reported ARR understates the demand for its product. If Anthropic had enough compute, he suggested, it might already be doing well north of $100 billion in ARR — perhaps $150 billion or $200 billion. On that basis, he proposed thinking about valuation against “unconstrained run-rate revenue,” a phrase he coined in the exchange to describe demand that exists but cannot be served because compute is scarce.

That scarcity also shows up in product quality. Baker argued that Anthropic has “deprecated the intelligence of Claude” because of compute constraints, citing an analysis that Claude, even on Opus, was producing 70% fewer tokens for the same question. Token quantity, in his view, is connected to answer quality and “quality of thinking,” even if intelligence density per token also matters. He said he had felt that degradation as a user.

O’Shaughnessy asked why Anthropic or OpenAI do not simply raise $100 billion at much higher valuations if investor demand is so intense. Baker’s answer was that capital discipline can itself become a strategic asset. The business is capital-intensive, the world is uncertain, and access to future capital and compute matters. Elon Musk was his model: Musk, Baker said, has treated making investors money as a “sacred covenant,” and because he has done that for 20 years, he can raise capital when he wants. The method, in Baker’s description, was simple: do not be greedy on valuation.

The infrastructure argument begins with “watts and wafers.” Gavin Baker is more confident capitalism will solve the watts shortage than the wafer shortage.

On power, he said the binding constraint has already shifted in some places. A senior data-center infrastructure investor at a large private equity firm told him that energy and chips used to be the biggest gating factors; now zoning and approval are more important. Baker acknowledged real industrial constraints, including turbine capacity and the difficulty of producing large turbine blades, but his base case is that capitalism is good at solving such bottlenecks over time. Absent major regulatory or political blowback, he expects the watts shortage to begin easing in 2027 or 2028.

He also argued that geopolitical energy shocks can improve America’s relative industrial position. Asked about the Strait of Hormuz, Baker said its closure could be “relatively awesome for America” in the context of the current administration’s goals. His reasoning was relative: the key input into U.S. electricity prices, natural gas, had fallen 20%, while natural gas in Asia and Europe had doubled or tripled. Since electricity is a major manufacturing input and directly feeds AI infrastructure, America’s relative manufacturing competitiveness improved overnight.

Wafers are harder. Baker described TSMC as a small group of “flinty older humans in Taiwan” who are among the most important people in the world economy. In his telling, TSMC represents an overwhelming share of Taiwan’s GDP, water usage, and electricity usage, and its leaders view themselves as inheritors of Morris Chang’s legacy. He recalled visiting Taiwan’s Science Park more than 20 years ago and asking whether TSMC could catch Intel. The answer he remembered was that catching Intel was a “beautiful dream” for their grandchildren. They eventually did it, partly because of Intel’s self-inflicted wounds, but also because TSMC thinks differently.

The wafer bottleneck may be what prevents an AI bubble. Baker invoked Carlota Perez’s work on foundational technologies and Michael Mauboussin’s idea of a “breakdown in diversity,” where market participants all become bullish on a new technology. Historically, foundational technologies have tended to produce bubbles: markets correctly identify the importance of the technology, speculative capital funds the build-out, supply gets ahead of demand, and then there is a crash. The crash is especially severe when the build-out is debt-funded, as Baker argued it was in 2000.

He stressed that the current cycle differs from 2000 in important ways. The AI build-out is still overwhelmingly funded out of operating cash flows rather than debt. Valuations are different. And unlike the fiber build-out, where he said 99% of fiber was unutilized, every GPU today is running at 100% utilization. But the historical pattern still matters: canal bubbles, railroad bubbles, and the internet bubble all suggest that foundational technologies can attract excess capital even when the underlying technology is real.

His most concrete bubble indicator is TSMC capacity. If TSMC expanded as much as Nvidia wanted, Baker said Nvidia might sell $2 trillion, $2.5 trillion, or even $3 trillion of GPUs in 2026 or 2027. At some level of supply, he believes that would likely become an overbuild. If there is no bubble, he joked, investors should “throw a party” for TSMC because it will have single-handedly prevented one.

The risk is that Intel or Samsung eventually breaks discipline. Baker expects one of them not to stay restrained. If either becomes a meaningful second source, it could force everyone else to respond. The key question, then, is whether TSMC can maintain its leading-edge advantage — which he described as roughly 9, 12, or 15 months — while expanding enough to keep rivals from taking large share, but not so much that wafer scarcity disappears. That is the Goldilocks zone.

Baker’s more aggressive infrastructure claims sit downstream of the same scarcity problem. If permitting, power, and wafer capacity remain binding constraints, he expects capital and engineering talent to pursue more radical solutions. Two of those are orbital compute and the Terafab.

He wants “orbital compute” understood as racks in space, not as a science-fiction terrestrial data center. When people hear “data centers in space,” he said, they imagine a Pentagon-sized building in orbit and dismiss the idea. That is the wrong frame. A Blackwell rack, in his description, weighs about 3,000 pounds and is roughly eight feet high, four feet deep, and three feet wide. The satellite is the rack.

The rack, as Baker described the concept, would have solar wings perhaps 500 feet long on each side, operate in a sun-synchronous orbit so the panels remain in sunlight, and use a radiator extending hundreds of feet behind it. Racks would be connected by lasers through vacuum, a capability Baker said already exists on Starlink satellites.

The objections are cooling and repair. On cooling, Baker said he has spent substantial time at Starbase and spoken with SpaceX engineers, whom he called “the most talented group of engineers on planet Earth.” In Baker’s account, the engineers he has spoken with are confident they have solved it. On repair, he was blunt: until there are “floating Optimuses,” failed orbital racks probably are not repaired. But if terrestrial regulation and permitting become the binding constraints, he thinks SpaceX could still sell as much orbital compute as it can make.

His case rests partly on what he says SpaceX already operates. Baker asserted that SpaceX runs the world’s largest satellite fleet, which he described as 98% or 99% of all satellites in orbit. Every Starlink satellite is already being cooled, he said. Starlink v3, in his account, is expected to operate at 20 kilowatts, compared with about 100 kilowatts for a Blackwell rack. Baker also asserted that the same company now operates the largest data center on Earth.

Orbital compute would not make terrestrial data centers irrelevant. Training will remain on Earth for a long time, while inference is more sensible for orbital compute. More broadly, he expects the system to consume as much compute as it can get. Still, companies tied to terrestrial power production and cooling should think carefully if they are expanding capacity just as orbital compute becomes more credible.

The Terafab is the wafer-side version of the same ambition. Baker described it as a SpaceX effort, with Tesla also involved, to build the world’s largest semiconductor fab in America. O’Shaughnessy noted that fabs typically have long lead times. Baker’s confidence rests less on normal fab construction timelines than on Musk’s ability, in his view, to attract talent, force supplier attention, and compress schedules.

In Baker’s account, the project has a partnership with Intel, which would provide access to decades of institutional knowledge that is only a few quarters behind the leading edge. He expects the A-teams from semiconductor equipment suppliers — ASML, KLA, Lam Research, Applied Materials, and others — to focus on the project. In his telling, one reason TSMC caught Intel was that the equipment suppliers wanted TSMC to catch up; they did not like having a monopsony, so their best people worked in Taiwan while Intel made mistakes.

Talent is the other pillar. Baker argued that Musk’s reputation in hardware engineering will attract elite engineers, especially from Taiwan, Japan, South Korea, and China. He imagined “Taiwan town,” “Japan town,” and “Korea town” near the Terafab, built around the restaurants, staff, and cultural details that would make engineers willing to move to Texas. That, he said, is not how Intel or Samsung think.

When pressed on timing, Baker answered with Musk’s operating history rather than a fab schedule. Others take three years to build a data center, he said; Musk built one in 122 days. Samsung, he added, had to give Musk an office in its Texas fab because he was unhappy with the pace of expansion. Baker’s conclusion was not certainty. It was: “We’ll see.”

DeepSeek sharpened the question Gavin Baker considers central for model-layer investing: will frontier tokens continue to capture most of the economic value?

So far, the answer has been yes. Baker said the overwhelming share of model-layer returns has accrued to the frontier, and that this surprised him. His example was Gemini 1.5 Pro. When it launched, he found it mind-blowing. Today, he called it “intolerable.” That experience made him more open to the possibility that frontier models will keep commanding premium economics longer than he previously expected.

His preferred lens is the Pareto frontier of intelligence versus cost. Nine months earlier, he said, Google dominated that frontier at every point, with OpenAI, xAI, and Anthropic inside it. Now he sees the frontier dominated by Anthropic and OpenAI, with Grok 1.5 also on it as the best low-cost 500 billion parameter model, while Gemini 1.5 is “hanging on.” He attributed Google’s loss of per-token cost leadership to conservative design choices in TPU v5 and Nvidia’s continued willingness to make aggressive choices.

Usage-based pricing reinforces the frontier’s advantage. Baker said that in earlier conversations he told investors to pay for the $250-a-month AI plans to develop an intuitive sense of frontier capability. He no longer thinks that works. To understand frontier AI today, even for non-coding uses, he said one needs Claude Code or Codex and an enterprise plan. Consumer and prosumer subscriptions are heavily rate-limited and, in his word, “lobotomized.”

That shift is bullish for model revenue. He compared it to telecom in 2005–2007, when cellular was a strong growth industry because users had fixed monthly plans with usage charges above included minutes. Growth slowed, he said, when the industry moved to all-you-can-eat pricing. AI is moving the other way: from all-you-can-eat to “pay-by-the-drink.” If users like using AI as much as they liked long-distance and cellular communication, and if one person can now run 100 agents, Baker expects usage-based enterprise models to drive enormous revenue. He predicted OpenAI and Anthropic could exceed $200 billion in ARR this year as more compute comes online and pricing shifts.

There is a distributional cost embedded in that forecast. Baker called it “sad for the world” that the best AI is increasingly available only to those who can afford usage-based frontier access.

The technical risks are just as important. Baker named three major questions: whether the “bitter lesson” continues to hold, whether frontier tokens keep their premium, and whether continual learning arrives soon.

By the bitter lesson, he meant the idea that more compute and data tend to outperform human algorithmic ingenuity. A violation of that principle is, for him, the biggest risk to the entire AI trade. People closest to AI, he said, are generally skeptical such a violation will occur. Baker is somewhat less skeptical because he thinks the world may be close to artificial superintelligence. If a 300- or 400-IQ system wants to become smarter and gain more resources, it may first make itself more efficient. That could create at least a temporary violation of the bitter lesson — not because humans out-invent scaling, but because AI systems do.

Continual learning is the other discontinuity. Baker defined it as a model dynamically updating its weights, or otherwise adjusting in real time, the way humans do. A human touches fire once and learns not to do it again. Today’s models, in his description, effectively need to put their hand in the fire a million times and then have designers include fire in the next training run or reinforcement-learning environment. There are crude forms today when something is verifiable, such as reinforcement learning during mid-training. True dynamic updating would be different. If it is solved, Baker said, “then we have a really fast take-off.”

Gavin Baker sees the proliferation of chip startups as healthy, including for Nvidia. But he is skeptical of companies trying to build “a better GPU.”

He explained chip design with an analogy from tank design. Tanks live inside an “iron triangle” of attack, defense, and mobility: more armor improves defense but reduces mobility. Chip designers face similar trade-offs imposed by physics and TSMC design rules. TPU, Trainium, and AMD are all, in his words, essentially trying to be better GPUs. Nobody is a better GPU, he said, though he thinks Trainium is currently doing best among the challengers.

Trainium 3 matters to him because it has a switch scale-up network, which he said is needed to economically inference mixture-of-experts models. Google, by contrast, had been in a torus architecture. AMD’s MI450 remains uncertain in his view.

For startups, Baker’s rule of thumb is that 1% market share in AI chips could be worth $100 billion. That is an excellent venture outcome, but the path is not to copy Nvidia. If a startup makes a different trade-off that is not hard, Nvidia can make the same trade-off, get better TSMC pricing, and use its model-company relationships to optimize faster. He warned venture investors against believing a startup has privileged access to a TSMC process that Nvidia has not already seen. “Jensen saw that process when it was a twinkle in Taiwan Semi’s eyes,” he said.

The opportunity comes from the disaggregation of prefill and decode. Patrick O'Shaughnessy summarized prefill as taking in the context and decode as writing the output. Baker added a colleague’s analogy: prefill is loading the cannon; decode is firing it. Prefill is the model understanding the prompt and tracking its own answer, and is fundamentally memory-capacity bound. Decode is generating new tokens, and is memory-bandwidth constrained. Those different workloads give chip designers a richer canvas: they can optimize aggressively for one or the other.

But the design still has to be hard to copy. Baker’s example was Cerebras, where his firm had been a venture investor. Cerebras chose wafer-scale computing, an architectural decision he called fundamentally different and difficult. It comes with trade-offs, including a ratio of on-chip compute and memory to shoreline I/O that creates issues when many chips need to be connected. Baker said Cerebras is working on ideas such as putting an optical wafer on top to address I/O constraints, and may be looking at hybrid bonding of DRAM. His broader point was that it took Cerebras three generations and roughly a decade to get the approach right. Startups need that level of resilience if they are going to survive failed first chips and iterate.

One unexpected consequence of the prefill/decode split, Baker argued, is longer useful lives for GPUs. Skeptics say companies are using unrealistic depreciation assumptions and that GPUs may only be useful for a year or two. Baker disagrees. If prefill and decode can be separated, older GPUs such as Hopper or even Ampere can be used for prefill, while systems such as Cerebras or Groq LPUs sit in front for other workloads. That could extend GPU lives to 10 or 15 years, “until it melts.” He noted that GPUs do melt, so there is still a physical limit, but said the innovation could lower financing costs and help private credit finance the AI build-out.

Gavin Baker’s venture framework outside infrastructure is the same: a company needs to be different, hard, and not obvious to the world before it reaches scale. If an idea becomes obvious before the company has scale, and execution is not hard, the company is in trouble.

Amazon was his counterexample. E-commerce was obvious to many people, but not to incumbent retail CEOs. Amazon also destroyed VC-backed e-commerce competitors by taking category margins deeply negative. Wayfair survived, in Baker’s telling, because its leaders built something operationally hard enough that Amazon tried to kill it and failed.

In AI applications, he thinks many founders are struggling with that test. In Jensen Huang’s “five-layer cake” of AI, as Baker described it, profits are accruing to energy, data centers, chips, and models — not much to applications. Cursor and Cognition are exceptions because they reached scale by focusing on code. Eighteen months earlier, Baker said, the companies truly focused on coding were Cursor, Cognition, and Anthropic, while OpenAI was doing everything. He thinks that focus was correct because coding may be the shortest path to useful AI or even ASI: if a model can code well, it can write tools to do many other things.

For narrower application companies, the question is whether they can reach a data moat before model companies enter the niche, or whether the niche is small enough that frontier labs will not bother but large enough to produce a venture outcome. Baker is skeptical of simplistic vertical-AI claims because the proprietary data in many niches is generated by humans. The bet is that a company can use that data to train a lower-cost model that frontier labs will not match. Maybe that works, he said, but investors have to be careful.

He connected this to the “token path,” a phrase he attributed to Jamin Ball at Altimeter. If a software or AI company is in the token path, it participates directly in AI usage. Databricks is in that path. If a company is not in the token path and is not doing something deeply niche, life may be hard.

There is a reversal scenario. If frontier-token premiums fall relative to other tokens, Baker expects an explosion of value creation at the application layer. The reason applications are struggling today is that the frontier layer captures so much of the value. If that changes, the distribution of economics could change quickly.

Open source is part of the same game theory. Baker believes Nvidia could get close to the frontier with its own model when it wants, citing Nemotron as an example of work in that direction. He does not think Huang wants to commoditize his complement, but he sees it as the logical counter-move to model companies trying to reduce Nvidia’s power.

His claims about open-source frontier models were sharper. Baker described current open-source frontier models as largely Chinese models trained with “stolen American tokens,” by which he meant distilled outputs from American frontier APIs. American labs, he said, are working hard on anti-distillation technology. The new prisoner’s dilemma is whether frontier labs continue to release their best models through APIs. If everyone at the frontier agreed not to release, Chinese open source would fall behind quickly, in Baker’s view. But if one company defects and releases, it will have the best model, revenue and cash flow, and more resources, which translate into intelligence. That would pressure everyone else to release as well.

Baker also rejected the idea that open source is free. Open-source tokens still require energy, GPUs, and often a revenue share for the model companies.

Gavin Baker called the current moment “the most exciting, thrilling time to be an investor,” but he is starting to worry about a diversity breakdown. He does not know anyone like him who is not bullish on DRAM, and he wishes there were more AI bears and more memory bears.

His concern is not that every AI valuation is high. It is that cross-sectional valuations do not cohere. He pointed to semiconductor capital equipment companies trading at roughly 40 times next quarter’s annualized earnings while DRAM companies trade at mid-single-digit multiples. At the peak of the last cycle, he said, the gap was more like five versus 12; at one point recently it was closer to three versus 45. Semiconductor equipment business models may have improved more than memory business models, and high-bandwidth memory may improve memory economics, but Baker argued the difference does not justify a 1,000% multiple gap.

He made a similar point about Nvidia and power-infrastructure valuations. Nvidia, he said, was in early April about as cheap relative to the market as it has been in 10 or 12 years, and cheap on an absolute basis. He found that hard to reconcile with a valuation like GE Vernova’s, because the latter seemed to build in an “unfathomable” amount of Nvidia share loss.

Shortages also distort which stocks work. Baker compared the dynamic to commodity bull markets. In a true commodity boom, the highest-cost suppliers often rise the most because they move from near bankruptcy to generating cash. That is happening across AI-adjacent industries, he said. Low-quality suppliers that buyers dislike because they are expensive, unreliable, or have high failure rates are nevertheless sold out and raising prices. Retail accounts on X and Reddit can then bid those stocks up dramatically. Baker finds that frightening because he expects many such names to go back down, especially if the companies do not allocate their windfall cash intelligently.

At the same time, the highest-quality AI companies are not obviously extended in his view, which makes him less worried. He contrasted real AI demand with what he called “nuclear” and “quantum” bubbles visible in 2024 and 2025. Some of that speculative behavior, he said, may now have spread into lower-quality AI-adjacent small caps.

Market structure has changed too. In 2024 and 2025, Baker said, the AI trade moved together. Investors could be long GPU compute, scale-up networking, and optical scale-across, and short power, and the trade worked from a risk-management standpoint because the factors held together. That broke in January. Scale-up networking could rise while scale-out fell; DRAM could underperform NAND and hard drives; cross-sectional correlations inside AI broke down. Investors had to become more granular.

Patrick O'Shaughnessy suggested AI itself may be creating price efficiency. Baker agreed. Better AI tools make it easier for more people to get smart quickly on narrow subsectors, trade them, and form baskets. That creates opportunities in names miscategorized by the market. His example was Astera, which he said had been placed in “copper loser” baskets even though its biggest product is expected to be a switch. Since switches and accelerators sit on the other side of copper or optical connections, Baker argued, a switch company cannot definitionally be a copper loser.

Asked to assess the major public technology companies, Gavin Baker drew distinctions around compute position, model progress, and willingness to engage with startups.

Google remains strong, in Baker’s view, because it has more compute than anyone, the largest installed base, valuable data, and YouTube data that may matter in robotics. Search and GCP add to the position. The caveat is that Google’s TPU advantage has faded. Baker was watching whether its next announcements would leapfrog OpenAI or Claude even slightly; if not, he said, that would suggest the Nvidia effect is even more powerful than he had thought.

Meta’s improvement is about organizational change. Baker gave Mark Zuckerberg credit for making the company AI-first internally, paying aggressively for AI talent, and producing Muse, which he called a major upside surprise. It is not on the frontier with xAI, Google’s entrant, OpenAI, and Claude, but Baker said it is close enough to be impressive. Meta’s absolute position is not as strong as Google’s, but its rate of change has improved.

Amazon looks strong because of Trainium. Baker also expects robotics to create real P&L efficiencies in Amazon’s retail business over the next 18 months. Its Nova models, he said, are not where Meta’s Muse is, but are better than they get credit for.

Microsoft drew the most nuanced treatment. Baker called Satya Nadella brilliant and exceptional, but said Microsoft “flinched” briefly in early 2025 when its capital-expenditure return algorithm seemed off. In AI infrastructure, if a company flinches, it can lose allocations and position that are hard to regain.

Now he thinks Nadella is taking the right risk. Microsoft could likely be growing Azure faster, and its stock could be much higher, if it used its GPUs simply to serve OpenAI and Anthropic demand. Instead, Nadella is using compute internally to improve Microsoft’s products and train its own models. Baker is skeptical Microsoft has the right team to succeed at frontier-model training, but he framed the decision as preparation for a world where frontier models may no longer be API-accessible. In that world, Microsoft needs its own capabilities.

Another distinction is startup engagement. Baker said Nvidia and Amazon are the two companies most deeply engaged with startups, by far. Google is next. Broadcom is engaged differently, as the favored ASIC supplier; for a startup, working with Broadcom on a second-generation chip is a level-up, and working with Broadcom on a first-generation chip is “manna from heaven.” By contrast, Baker said he sees essentially zero startup engagement from AMD, Microsoft, and Meta, with “a little” qualification. He thinks that could become a real disadvantage because some of the best teams are now at startups rather than inside large public companies.

Gavin Baker’s response to increasingly capable AI is not to search for what humans can still do unchanged. He used The Last Samurai as his metaphor. In his summary of the film, samurai courage and skill eventually meet the machine gun. “If we do not all become masters of the machine gun,” he said, “we’re gonna get massacred.”

For investing, that means integrating AI agents into the process. Baker said he now has agents running all the time. His most useful agent summarizes the points that would be interesting to him from podcasts. He feels professionally compelled to follow hours of material from OpenAI, xAI, Google, Cursor, Fireworks, Baseten, Jensen Huang, Elon Musk, Dario Amodei, and others, but does not have enough time. Agents help surface needles in haystacks.

He also uses AI for labor-intensive investment work such as reviewing management compensation. Baker cares whether executives are incentivized through simple RSUs or through PSUs, and what those PSUs reward. Pulling proxy details, comparing changes over time, and identifying compensation signals is the sort of task he thinks AI can do well, freeing people for more creative work.

The defensive side is cybersecurity. Baker said everyone needs a family or company “safe word,” created away from digital devices, because AI-enabled impersonation will make social engineering far more dangerous. He imagined an accurate FaceTime simulation of a child, parent, or grandparent that knows everything available from digital traces and asks for a wire transfer. The safe word, in his view, is a simple defense against that class of fraud.

Gavin Baker’s final concerns moved beyond public equities and infrastructure into personal safety, geopolitics, and social distribution.

He is increasingly worried about personal safety for public figures associated with AI, especially as AI becomes politicized amid what he described as a broader rise in political violence in America. Baker pointed to what he described as someone throwing Molotov cocktails at Sam Altman’s house as the kind of incident he fears could become more common.

At the geopolitical level, Baker argued that AI superiority can be destabilizing. Ukraine is “really starting to win,” he said, and in his view the reason is not only better drones, though he thinks Ukraine has those too, but battlefield AI. He described Ukraine as having the best battlefield AI outside probably America and Israel. As China and other adversaries process the military implications of U.S. AI advantage, he asked, how will they respond?

His optimistic countercase was another Pax Americana. Baker argued that after 1945, the United States had the nuclear bomb when no one else did and could have tried to control the world. Instead, it rebuilt Germany and Japan, demilitarized, and helped create a period of global stability. He suggested the West’s record is not told forcefully enough, including the British Empire’s role in ending slavery and America’s postwar restraint. In the AI context, the point was that overwhelming technological advantage does not have to produce domination or global instability; Baker thinks it could also produce a new period of American-led stability.

That optimism coexists with a strong claim that AI is an event horizon. Baker called himself an “AI optimist maximalist” and gave a concrete example: someone whose daughter had a rare mutation with no cure assembled resources, received compute from labs, spun up agents, identified an existing drug that could affect the disease, and created a company to pursue a cure. Her life, Baker said, is already immeasurably different because of AI.

But he ended on humility rather than triumph. The Luddites, in his view, will be wrong, but their concerns need to be addressed thoughtfully. Society needs to make sure AI is good for everyone. And he returned to the distributional problem he had raised earlier: it is dystopian if the best AI is available only to people with a lot of money. That, he said, needs to be solved.