Gavin Baker - Watts and Wafers - [Invest Like the Best, EP.473]
What this covers
This is the sixth conversation between the host and Gavin Baker, who runs Atreides Management. It opens on how fast Anthropic is growing, with Baker noting it added $11 billion of ARR in a single month, roughly the combined businesses of Palantir, Snowflake, and Databricks. From there it moves to power and orbital compute. Baker argues that orbital data centers only work economically with fully reusable rockets, which is why SpaceX has no near-term match. He also walks through Starlink's cooling and power figures, from about 20 kilowatts per V3 satellite up toward the 100-kilowatt scale of a Blackwell rack.
The middle of the conversation is about chips. Baker explains how splitting inference into prefill, bound by memory capacity, and decode, bound by memory bandwidth, opens room for new designs. It could also keep older Hopper and Ampere GPUs useful for 10 to 15 years, which he sets against skeptics who assume lifespans of one to two years. For chip startups he applies an Iron Triangle of physical trade-offs. He says a startup must pick a different point in that space that is also hard to copy, because Nvidia will match any easy move. He also offers a rule of thumb that 1% market share is worth about $100 billion.
The rest turns to markets and risk. Baker says he is starting to see a diversity breakdown as skeptics capitulate, while arguing the buildout differs from 2000 because it is funded largely from cash flow and every GPU is in use. He names a violation of the Bitter Lesson as the biggest risk to the AI trade. He also covers distillation by Chinese open-source models, a prisoner's dilemma over releasing frontier models through APIs, the shift to pay-by-the-drink pricing, and why TSMC's capacity choices are the one thing to watch.
Baker argues that AI's near-term growth is constrained by physical inputs, power and advanced wafers, that the frontier-model layer keeps capturing most model economics, and that TSMC's capacity discipline is the key variable deciding whether the buildout avoids a 2000-style bubble.
- Zoning, approvals, and turbine capacity are now bottlenecks, with power shortages easing in 2027-28 and orbital compute addressing the longer term
- The overwhelming majority of model-layer economic returns sit at the frontier, and the Pareto frontier is shifting toward Anthropic and OpenAI
- The current buildout is funded by operating cash flow at 100% GPU utilization, unlike the debt-fueled 2000 buildout, but supply discipline at TSMC is what prevents overbuild
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
A violation of Sutton's bitter lesson, in which human algorithmic ingenuity beats more compute, is the biggest risk to the AI trade, and near-ASI systems may make a temporary violation possible because they can make themselves more efficient.
“A violation of Richard Sutton's bitter lesson is for sure the biggest risk to this trade, to all of AI.”
Chip design, like tank design, faces an Iron Triangle of trade-offs imposed by the laws of physics and foundry design rules, so new chip entrants must choose a different point in that trade-off space rather than simply making a better GPU.
“Chip design is the same. There are these fundamental constraints imposed by the laws of physics”
Based on prior technology precedents, a foundational technology like AI gets a diversity breakdown in which everyone becomes bullish, which funds a bubble; supply then outruns demand and a crash follows, which is particularly severe when the buildout is debt-funded as in 2000.
“There's what Mauboussin calls a breakdown in diversity. Everyone becomes bullish on this new technology.”
Valuation-sensitive investors massively underperform during bubbles and get fired by clients, as George Vanderheiden did when he fought the 1999 bubble and retired in early 2000, even though he was right.
“if you're valuation sensitive, you massively underperform. You get fired by probably all your clients.”
In commodity shortages, the highest-cost, lowest-quality suppliers go up the most because they go from near-bankruptcy to gushing cash, which makes commodity investing hard since quality outperforms over the full cycle but the high-cost names win during the upswing.
“In a real bull market for a commodity, the commodity suppliers with the highest costs go up the most”
Baker is beginning to worry about a diversity breakdown, a Mauboussin-style state in which nearly everyone becomes bullish on a new technology, because that consensus is the precursor to a bubble, and he sees more capitulation among previously skeptical investors.
“There's what Mauboussin calls a breakdown in diversity. Everyone becomes bullish on this new technology.”
Disaggregating inference lets older GPUs like Hopper and Ampere handle prefill behind specialized decode chips, extending their useful life to 10 or 15 years, which counters skeptics who assume 1 to 2 year lifespans and supports GPU-backed private credit.
“The disaggregation of inference means that these GPUs are going to have 10 or 15-year lives.”
If AI achieves continual learning, meaning it dynamically updates its weights in real time, there would be a very fast takeoff, because current models need a million trial exposures and training-run patches while humans learn from one hot-stove experience.
“If we get that, then we have a really fast takeoff.”
An overwhelming majority of economic returns at the model layer have accrued to frontier models, which is surprising, and whether frontier tokens keep capturing that value is one of the most important questions an investor must hypothesize about.
“All the economic returns to AI at the model layer — not all of them, but an overwhelming amount of them have been at the frontier”
Chinese open-source models are advancing largely by distilling American frontier tokens, which the American labs are trying to stop with anti-distillation technology, and open-source models still get monetized through GPU margins and revenue shares.
“which today consists of Chinese models with stolen American tokens.”
DeepSeek was the first time it became clear how much more compute reasoning models consume during inference than non-reasoning models.
“This was the first time we saw how much more compute-hungry reasoning models are during inference than non-reasoning models.”
Disaggregating prefill from decode opens new chip design trade-offs: prefill, which ingests the prompt and tracks state, is bound by memory capacity, while decode, which generates new tokens, is bound by memory bandwidth.
“That is fundamentally a memory-capacity-bound problem. Decode is the process of generating new tokens, and that is memory bandwidth constrained”
Orbital compute only makes economic sense with fully reusable rockets, and SpaceX's consistent reusable landings are the reason no other company can yet match the economics.
“None of this makes sense without reusability, and that means you have to land it.”
Anthropic added $11 billion of ARR in a single month, roughly equal to the combined businesses of Palantir, Snowflake, and Databricks, which took about 10 years to build.
“Anthropic added $11 billion of ARR”
Starlink satellites are already cooled in orbit, and Starlink V3 is designed for about 20 kilowatts, with SpaceX likely scaling toward 60 kilowatts and then 100 to 120 kilowatts, which is comparable to a 100-kilowatt Blackwell rack.
“Every Starlink — they're cooling it today. Starlink V3 is going to operate at 20 kilowatts.”
The current buildout is still overwhelmingly funded out of operating cash flow and is valued more reasonably, and every GPU runs at near 100% utilization, whereas about 99% of fiber was unutilized in 2000, so the fundamentals differ from the year 2000 bubble.
“it's still overwhelmingly funded out of operating cash flows”
A new prisoner's dilemma exists for frontier labs over whether to release models via API, since if all agree not to release, a Chinese open-source defector would gain the best model, revenue, and cash flow, and then others would be forced to release.
“The new prisoner's dilemma is going to be, if you are at the frontier, do you release that model via API or not?”
TSMC's constraint on wafer supply can prevent an AI bubble because, if TSMC had followed Nvidia's demand, Nvidia could sell $2 to $3 trillion of GPUs, which would overbuild the market; so if no bubble occurs, TSMC deserves credit for having prevented it.
“this fundamental shortage of wafers, which today is controlled by Taiwan Semi, will prevent one.”
In venture, if an idea becomes obvious to the world before a company builds scale, the company is in trouble, because scale is the ultimate advantage, so investors should ask whether an idea is non-obvious, different, and really hard to do.
“if it's not hard to do, if it becomes obvious to the world before you have built scale — scale is the ultimate advantage — you're in trouble.”
If one thing were watched to judge whether there is a bubble, it would be TSMC's capacity decisions, because wafer supply is the constraint that can prevent overbuild.
“If I were to watch one thing to understand whether there's a bubble, it's Taiwan Semi's capacity decisions.”
AI is shifting from all-you-can-eat subscriptions to pay-by-the-drink usage pricing, the way cellular and long-distance once did, which means frontier capability now requires enterprise plans and makes it hard for individual investors to build intuition about what AI can do.
“AI is just shifting from all-you-can-eat to pay-by-the-drink.”
A rule of thumb for chip startups is that 1% market share is worth about $100 billion, a good venture outcome, but startups must do something different that is also hard to do, since Nvidia will copy any easy trade-off.
“my rule of thumb is, 1% market share is going to be worth $100 billion.”
The DeepSeek selloff was a setup that careful observers could read in advance: the paper appeared 7 days before DeepSeek Monday, Asian AWS GPU prices had already doubled, and GPU availability was falling, so the implication for compute demand was visible before the market reacted.
“Prices in the AWS availability zones in Asia had already doubled. You were seeing GPU availability go down.”
Nvidia gets better foundry prices than any startup and works with every model company, so if a chip startup makes an easy trade-off to optimize memory capacity, Nvidia will make the same trade-off, and startups should not expect to win by doing what is merely obvious.
“Nvidia is going to make those same trade-offs. They get better prices from Taiwan Semi than you're ever going to get.”
Software and AI companies must be in the token path, meaning they must be part of the flow of tokens through the model stack, and companies outside the token path in non-niche areas face hard prospects.
“If you're a software company or an AI company of any kind, you have to be in the token path.”
Trillions of dollars of value has been destroyed at the AI application layer, even counting the most successful natives like Cursor and Cognition, and the companies gaining value are those with the highest ratio of utilized GPUs per human.
“trillions of dollars of value has been destroyed by AI at the application layer”
Google lost its per-token cost leadership because of very conservative design decisions for TPU v8, which were meant to take share partially away from Broadcom, while Nvidia continued making aggressive design choices.
“Google losing their per-token cost leadership as a result of making very conservative design decisions with TPU v8”
Claude now produces about 70% fewer tokens for the same question, even on Opus, which Baker reads as a deprecation of intelligence, since token quantity correlates with answer quality and thinking depth, and intelligence density per token also matters.
“Claude is generating, even on Opus, 70% less tokens for the exact same question.”
The ideal is a Goldilocks zone in which TSMC expands enough to keep Intel or Samsung from becoming an at-scale second source with over 30% market share, while still keeping wafer supply constrained enough to avoid a bubble.
“There's a Goldilocks zone where they expand enough to make it hard for Intel or Samsung”
Elon's reputation in hardware engineering will attract A-teams from semiconductor equipment companies and top talent to Terafab, mirroring how equipment makers helped TSMC catch up because they did not want a monopsony.
“They don't like having a monopsony. The A-teams were in Taiwan working.”
Google's TurboQuant memory optimization, published a year earlier and publicized on X during Google's negotiation of a long-term agreement with Micron, Samsung, and SK Hynix, triggered a DRAM selloff, though Baker could not find an AI engineer who believed it would affect DRAM demand.
“I was unable to find a single AI engineer on planet Earth who believed that TurboQuant would have any impact on DRAM demand.”
Cerebras chose a hard, fundamentally different architecture by doing wafer-scale computing, which took three generations of chips to get right, and it could solve I/O shoreline problems through optical wafers or hybrid bonding of DRAM.
“It took Cerebras three generations of chips to get it right.”
Amazon and Nvidia are by far the most deeply engaged with AI startups, Google is next, while AMD, Microsoft, and Meta show essentially zero startup engagement, which may give Nvidia and Amazon a lasting advantage in identifying the best teams.
“The two companies who are the most deeply engaged with startups are Amazon and Nvidia by a mile.”
Microsoft would likely be an $800 stock if it used its GPUs to serve OpenAI and Anthropic capacity instead of its own products, and Satya Nadella's decision to use compute internally is risky but correct for a world where frontier models are not API-accessible.
“Microsoft probably would be an $800 stock today if they were using their GPUs to serve solely OpenAI and Anthropic's capacity”
The head of data center infrastructure investing at a large PE firm reported that the gating factor has shifted from energy and chips to zoning and approvals, which are now much more important constraints on AI buildout.
“Now it's zoning and approvals — much more important.”
Gavin is worried that people who were skeptical of AI a year ago are no longer skeptical, a sign of diversity breakdown, while high-quality AI companies remain unextended in valuation.
“it worries me a little bit that people who were very skeptical a year ago are no longer skeptical.”
Elon Musk's consistent focus on making investors money, and never pushing valuation, has given him a superpower to raise as much capital as he wants whenever he wants, with benefits that can last 20 to 30 years.
“He can essentially raise as much capital as he wants whenever he wants.”
Intel and Samsung will not stay disciplined on capacity, and when one breaks, it will force everyone else to break too, so the pace at which TSMC maintains its lead over them matters most.
“They will break, and then at some level that will force everyone else to break.”
Management incentives can be read from compensation design, since RSUs versus PSUs and their metrics reveal what executives are incentivized to do, and AI systems can now produce a first-pass analysis of proxy filings that frees analysts for creative work.
“I'm very sensitive to management compensation. What are they incentivized to do? Do they just have stupid RSUs, or do they have PSUs?”
Because racks in space are linked by lasers through vacuum, the rack can be physically bigger, so the binding constraint is weight rather than size, unlike terrestrial data centers where copper cabling forces racks to be compact.
“you can make the rack bigger physically. You're focused on weight, not size.”
Cross-sectional correlations within AI fell apart in January 2026, so scale-up networking rose while scale-out fell and DRAM massively underperformed NAND and HDDs, requiring much finer-grained hedging than the AI basket trade allowed.
“That all blew out in January of this year.”
Jensen Huang could probably get close to the frontier with his own model whenever he wants, and Nvidia's logic is to keep open source a timeframe behind the frontier as a countermove to OpenAI and Anthropic.
“Jensen can probably get pretty close to the frontier.”
There are two kinds of drawdowns: those where your thesis is wrong and you must crystallize the loss, and those where you underperform in companies you know deeply and disagree with the price action, which lets you build pent-up future alpha by leaning in.
“there are drawdowns where you're wrong... And then there are drawdowns or periods of underperformance where you're underperforming because of companies you know really, really well”
The closure of the Strait of Hormuz was relatively favorable for America: US natural gas (NG1) fell about 20% while gas in Asia and Europe doubled or tripled, improving US relative manufacturing competitiveness overnight, which made it easier to stay focused on AI fundamentals.
“that was down 20%. While natural gas in Asia, Europe, and everywhere else doubled or tripled.”
Among custom accelerators such as TPU, Trainium, and AMD, Trainium is probably doing best today, and Trainium 3 needs to ramp because its switch scale-up network is required to inference mixture-of-experts models economically.
“Today, I think probably Trainium is doing the best.”
AI coding companies like Cursor and Cognition reached scale by focusing on code, because coding may be the shortest path to ASI and useful AI since a model good at coding can write code to do anything.
“coding might be the shortest path to ASI and useful AI. Because if you're really good at coding, you can write yourself code to do anything.”
Everybody needs a family or company safe word that cannot be socially engineered, because AI can simulate a loved one's voice and face convincingly enough to trick someone into wiring money.
“everybody needs to have a safe word”
Terafab is a SpaceX and Tesla joint venture to build the world's largest fab in America, partnering with Intel, which gives it access to decades of institutional knowledge and a position 3 to 5 quarters behind the leading edge.
“It's a SpaceX — I believe Tesla's involved as well — joint venture to build the world's largest fab, here in America.”
Inference is well suited to orbital compute while training will remain on Earth for a long time, so terrestrial data centers stay valuable for Baker's lifetime even as orbital compute grows.
“Inference is very sensible for orbital compute. Training will be done on Earth for a long time.”
Private credit is in pain from SaaS loans, which probably need to be marked down further, since debt-laden software companies struggle to adapt to a very different margin structure when public SaaS companies are already struggling.
“private credit is in pain from these SaaS loans, and however much they're marked down, they probably need to be marked down more.”
To understand what frontier AI can do even for non-coding use cases, an investor now needs Claude Code or Codex on an enterprise plan, because cheaper flat plans rate-limit users to degraded outputs.
“To understand what frontier AI is capable of today — even for a non-coding use case — you need to have Claude Code or Codex, and you need to be on an enterprise plan.”
Anthropic has a dramatically lower cost per token than OpenAI, having burned roughly 80% less cash to reach a similar revenue scale, so the two companies have very different structural returns on invested capital.
“I think they've burned maybe 80% less than OpenAI.”
Orbital compute should be understood as individual racks in space connected by lasers into a virtual data center, not as a Pentagon-sized floating building; a Blackwell rack weighs about 3,000 pounds and the satellite is roughly rack-sized with solar wings hundreds of feet long.
“I would like to redefine orbital compute as racks in space”
Elon Musk built a data center in 122 days, while others take about 3 years, and Samsung gave him an office in its Texas fab because he was unhappy with its build pace.
“Everybody else has taken 3 years to build a data center; he built one in 122 days.”
If GPUs can be financed at 5 to 6% rather than the low-7% rates CoreWeave obtained, the cost of financing the AI buildout changes mathematically, so the technological extension of GPU lives lowers the cost of capital for compute.
“if you can start to finance GPUs at more like 5% or 6% instead of”
Some of the best outside-quality opportunities are in miscategorized names such as Astera Labs, which was lumped into copper-loser baskets even though its switch products use both copper and optics, so switch and accelerator companies cannot be copper losers.
“Astera was in a lot of copper-loser baskets.”
Cross-sectional AI valuations do not make sense: semiconductor capex companies trade near 40 times next quarter's annualized earnings while DRAM companies trade at mid-single-digit multiples, and GE Vernova's valuation implies an unfathomable amount of Nvidia share loss.
“semi-cap equipment companies trading at 40 times next quarter's annualized earnings and DRAM companies trading at mid-single digits.”
Capitalism will solve the watts shortage over time, absent major regulatory or political blowback, with the power shortage beginning to ease in 2027 and 2028 as alternative energy sources and turbine capacity come online.
“So the watts shortage will probably begin to alleviate in '27, '28”
Owners of large installed bases of scarce compute, such as the biggest CPU and GPU fleets at hyperscalers, benefit from shortage, and CPUs are becoming more important in an agentic world because they handle orchestration and tool calls.
“if you own a giant installed base of what is currently in shortage, that's also a very good place to be.”
Zuckerberg made Meta an AI-first company internally, uniquely among internet giants, and Meta's Muse model from MSL is close to the Pareto frontier, so Meta is in a better position despite less absolute advantage than Google.
“Muse was a really big upside surprise.”
Amazon is in a strong position because of Trainium, and it should see real P&L efficiencies from robotics in its retail business over the next 18 months, while its internal Nova models are better than they get credit for.
“Amazon is in a really strong position because of Trainium. You're going to see real P&L efficiencies from robotics over the next 18 months”
If Anthropic could get all the compute it wanted, it would likely be doing well north of $100 billion in run-rate revenue, possibly $150 billion, so buying it today is roughly 5 times unconstrained run-rate revenue.
“they'd probably be doing well north of $100 billion today. Maybe $150 billion.”
Companies ramping power and cooling capacity for Earth data centers risk being caught with excess capacity just as skeptics come to accept that orbital compute is real.
“a lot of these capacity ramps are going to be hitting just as all of the silly skeptics start to understand that orbital compute is very real.”
The shift to usage-based enterprise pricing means OpenAI and Anthropic could exceed well over $200 billion in ARR this year, and they can push frontier token pricing as more compute comes online.
“the shift to usage-based pricing is probably why you will see OpenAI and Anthropic exceed well over $200 billion in ARR this year.”
Google dominated the intelligence-versus-cost Pareto frontier nine months ago, but now Anthropic and OpenAI dominate it, Grok 4.3 sits on it as the lowest-cost large model, and Gemini 3.1 is hanging on, possibly subsidized out of pride.
“Google dominated that 9 months ago.”
On a relative basis, tech was about as cheap versus the rest of the market as at any point in the last 10 years, as of early April 2026, despite AI being in an extraordinary growth moment.
“tech essentially got as cheap as it's been versus the rest of the market”
Investors should expect a bubble based on 200 to 300 years of history, including railroad, canal, and South Sea bubbles, even though no one wants one.
“we should expect a bubble. That's terrifying. Nobody wants a bubble.”
Ukraine is winning largely because of battlefield AI, and as the US advances its AI edge, adversaries like China may respond in destabilizing ways, though a Pax Americana informed by AI dominance might stabilize the world instead.
“The reason Ukraine is really winning is they have the best battlefield AI”
Post-1945 American restraint, which did not exploit its nuclear monopoly, created Pax Americana, and AI dominance could similarly produce a stable new American-led order rather than destabilization.
“Maybe it leads to another Pax Americana informed by our AI dominance.”
Harness engineering is not as important as the model itself, but it matters a lot, and the two are increasingly co-developed; the harness acts as the runtime that knows tools, creates context and memory, tracks state, and supplies prompts.
“It turns out that harness engineering is not as important as the model, but it really matters.”
SpaceX compounded at a low-30% annual rate for a decade, which Baker attributes to Elon Musk's focus on preserving his capital-raising advantage and balancing investors and employees.
“SpaceX compounded at low 30% per year for a decade”
Political violence is rising in America as AI becomes more political, and AI leaders with large public profiles face growing personal safety risks, with Molotov cocktails thrown at Sam Altman's house as an example.
“someone threw Molotov cocktails at Sam Altman's house”
Musk is revered in China, Taiwan, South Korea, and Japan, and he will recruit the best engineers by moving restaurants and whole neighborhoods of staff from Taiwan and other countries to Texas near the Terafab, a strategy Intel and Samsung leaders do not think like.
“Next to the Terafab, there'll be a Taiwantown.”
Unlike the 1970s, the US economy is far less energy intensive and is now the world's largest oil and gas producer and exporter, so an energy shock is less likely to produce actual shortages.
“the US economy is dramatically less energy intensive than it was”
Cerebras has an overwhelming ratio of on-chip compute and memory relative to shoreline I/O, which limits scale-up networking, and it is exploring placing an optical wafer on top of the chip, while claims that it cannot run arbitrary model sizes are not true.
“I/O is bound by what's called the shoreline — the sides of the chip.”
Nvidia and TSMC have no formal contract, operating instead on handshake trust and perceived fairness, a relationship Baker finds remarkable given the scale of their dealings.
“It's wild that Jensen has never had a contract with Taiwan Semi.”
Taiwan's semiconductor leaders see themselves as inheritors of Morris Chang's legacy and the Silicon Shield, and they took a multi-decade view when told it was a dream for their grandchildren to catch Intel.
“They talk about the Silicon Shield. They all view themselves as inheritors of Morris Chang's sacred legacy.”
Using AI agents and compute from the labs, Gavin's family assembled resources that produced a drug on the market that can impact his daughter's rare disease, and then a company was spun up to cure it.
“came up — using AI — with a drug on the market that can actually impact his daughter's disease.”
In 2024 and 2025, people ignored obvious nuclear and quantum speculative bubbles while questioning whether AI was in a bubble, and that speculative spillover now reaches lower-quality small caps moved by X and Reddit.
“You have this nuclear bubble and this quantum bubble right here, right in front of you.”
Gavin's most useful AI agent summarizes the interesting points from roughly 6 hours a day of podcasts and talks by AI leaders, and AI now does a first pass on management compensation structures such as PSUs versus RSUs, freeing analysts for creative work.
“My single most useful agent is a really good summary of the points that would be interesting to me from podcasts.”
Many companies are waiting until after the midterms before taking action such as workforce reductions, because no one wants to be a piñata during the midterms.
“A lot of companies are waiting till after the midterms to take action in terms of maybe workforce reductions.”
The two main objections to orbital compute, radiator cooling and repairing a failed rack, are not yet solved; Baker believes SpaceX engineers are confident but that repair and maintenance remain a focus, and that Starship must mature for it to work.
“There's some engineering that needs to happen to turn Starship into a Mars Colonial Transporter.”
It is dystopian that the best AI is only available to people with a lot of money, so society must address this and approach the transition with humility, recognizing that the Luddites' concerns deserve careful responses even if they are ultimately wrong.
“It is a little dystopian that now the best AI is only available to people with a lot of money.”
Google lost its TPU advantage but remains well-positioned because it has the largest installed compute base, valuable YouTube data for robotics, and dominant search and cloud businesses, so it is unlikely to fall out of a strong position.
“they have the biggest installed base of compute.”
Vertical AI startups betting on proprietary niche data to train cheaper models than frontier labs are risky, because the data is generated by humans and frontier labs are skeptical that such niches can be defended.
“You're betting that you're able to use that proprietary data in this narrow vertical to train a model that's lower cost than the frontier labs can ever get to.”
The West has not produced the large blades needed for power turbines in about 80 years, and only two machines can cast them, a constraint that capitalism will eventually work through.
“We haven't made one in 80 years in the West.”
Musk has done more than any living human to decarbonize the world and is single-handedly bringing manufacturing back to America while reviving defense tech.
“He's done more than any living human to decarbonize the world.”
Anthropic could raise money at at least a 100% premium to its rumored latest valuation mark, and raising more capital is prudent given capital intensity and geopolitical uncertainty.
“But could Anthropic raise money at probably at least a 100% premium to this rumored latest mark? Of course.”
Humans should aim to master new AI tools, like the machine gun in The Last Samurai, so they can keep contributing value.
“So I am trying to become a master of the machine gun.”
The SaaS and cloud revolution created roughly $5 to $10 trillion of value, which Baker uses as the scale against which Anthropic's monthly ARR gain is compared.
“the SaaS and cloud revolution created, we'll call it, between $5 and $10 trillion of value”
The speed of AI revenue growth in 2026 is unprecedented in the history of capitalism, and he has never seen an exponential like it, exceeding even the DeepSeek episode in extremity.
“Nothing like that has ever happened in the history of capitalism”