Baker came up at Fidelity covering chips through the 2000 cycle and now runs Atreides. A wide-ranging conversation on where the AI build binds, the TSMC bottleneck, why AI's rise runs on silicon, and the first-principles case behind his Tesla and Nvidia conviction.
[Note: this was recorded back in late February 2026]
The short version, if you do not have the hour:
- Baker's read on the bubble question is precise, not reflexive. Tech trades at roughly the same multiple it did five or six years ago, multiples have compressed since the start of 2025, and tech now trades at a discount to consumer staples, which almost never happens. On valuation, he argues, there is no bubble.
- The live risk sits in capex rather than valuation. His test for whether an overbuild ends in a crash is whether it was funded with debt or with cash flows, and so far the ROI on AI spend has been positive.
- There is a temporary "divot" in that ROI because Blackwell is an enormous expenditure being used for training, which does not pay back immediately. Early agentic models he cites as GPT-5.2, Grok 4.20, and Codex 5.3 tell him the payback on Blackwell will be high.
- Two forces, he hopes, prevent a 2000-style overbuild: the world is fundamentally short "watts and wafers," and the scars of the dot-com crash (down 80-85% then, versus down 60-65% in 2009) put a permanent lid on tech valuations.
- The wafer bottleneck is TSMC. Baker likens chip manufacturing to baking, where the tools are shared ASML equipment and the edge is the recipe, so a leader is very hard to catch. He argues TSMC will not expand capacity fast enough to relieve the shortage.
- His Tesla conviction dates to a 2010 bus-tour meeting with Elon at a ~$1.5B market cap, and it rests on a deflationary-input argument: battery energy density has been compounding while every other car input inflates.
- On AI and video games, world models cut game-development cost by about 90%, good for platforms and hard on game makers, but games keep rendering locally on GPUs for the next five to seven years. Rendering a live game like Monopoly Go through a video model would cost more than 100x its revenue.
- His durable lesson is a Fidelity mentor's line: as an investor you either panic early or double down late, almost no one does both, and the job is knowing which one you are.
- Must watch: Episode 56: Gavin Baker - Managing Partner of Atreides Management
- Follow @firesidealpha for highlights of the best conversations in technology and business.
Gavin Baker has been answering the same question for most of his career, and the question is whether the thing in front of him is a bubble. He started at Fidelity in 1999 covering semiconductors as the dot-com bubble crested and broke, rose to run the flagship $17 billion OTC fund, and made Fidelity one of the largest early institutional holders of both Nvidia and Tesla. From 2007 through 2021, by his own count, he was asked about "the bubble" in ninety percent of client meetings, back when the bubble in question was Google and Meta and Amazon and Apple. Now the question is AI, the money at stake is far larger, and Baker's answer is more careful than either the bulls or the bears want it to be. Growing up in Texas with an unlimited book budget and a lifelong history habit gave him the frame he applies now, which is that markets have run this experiment many times before, and the pattern is legible if you are willing to read it.
The 2000 call
The reason Baker's bubble read carries weight is that he made the call at the actual top, six weeks into covering semiconductors as a 23-year-old, using the same supply-and-demand framework he had picked up starting his career in a commodity industry. In February 2000, he pulled the inventory data on every semiconductor company and their customers, and the picture was uniform.
- Customer inventories and days of inventory were at all-time highs, with finished goods at an all-time high.
- Semiconductor inventories themselves were at an all-time high.
- The stocks were trading at multiples no one had ever seen.
His note said the demand environment did not support those valuations, and it landed harder than he expected. A senior analyst he names as Rajeev Kaul printed 50 copies, downgraded his own stocks, and walked the young Baker around to see every portfolio manager that day. The two names Baker liked out of the wreckage were both small then, Integrated Circuit Systems and Nvidia, which is where his 25-year habit of following Nvidia began. He also met a young Jensen Huang in that stretch, someone he now calls one of the two or three most exceptional people he has ever met, and he tied the point to a moat that does not show up in a model: only a handful of semiconductor leaders, in his telling Jensen, Hock Tan, and Lisa Su, reliably retain their best engineers, and that talent retention is a durable edge. The takeaway for the present is that Baker is not a permabull rationalizing a rally. He has called a semiconductor top before, at the start of his career, on inventory data most people were not reading.
Every revolution gets a bubble
Baker's starting point is historical rather than emotional. He leans on Carlota Perez and her book "Technological Revolutions and Financial Capital," and on three or four centuries of market history, to argue that a genuinely revolutionary technology reliably produces a bubble because the market is doing something rational.
"Anytime you get a truly revolutionary new technology, you always get a bubble. Financial markets, they correctly get excited about the new technology. Financial markets are efficient most of the time. So they correctly identify it as being revolutionary and world transforming. And then you have what Michael Mauboussin calls a breakdown of diversity of opinion, and that's what fuels a bubble." Gavin Baker
The sequence he lays out is the same across canals, railroads, radio, PCs, and the internet: correct excitement, then a breakdown in the diversity of opinion, then an overbuild of the new technology, then a pause in demand, then a crash. The distinction he flags as decisive for AI is how the buildout gets funded.
- Overbuild funded out of debt is the dangerous kind, because the debt service does not pause when demand does.
- Overbuild funded out of cash flows is survivable, because the spender can throttle without a solvency event.
For someone extremely disciplined on valuation, he is candid that a bubble is not exciting, it is a nightmare, because it is the enemy of every long-term investor. That is the lens he brings to the present, and it is why his answer to "are we in a bubble" splits into two very different questions.
Not a valuation bubble
On the first question, valuation, Baker is blunt that the bubble talk is not supported by the multiples.
"Anyone who says we're in a valuation bubble is just not paying attention. Tech's at the same multiple it was at five or six years ago. Tech multiples have compressed since the beginning of '25. Tech is now at a discount to staples, which happens very rarely. We're not in a bubble from a valuation perspective." Gavin Baker
The comparison that gives that its weight is tech trading at a discount to consumer staples, a defensive, slow-growth group, which he says happens very rarely. Put the margins, revenue growth, and returns on invested capital of a good tech company into any other industry, he argues, and the stock would trade at a large premium. So the question worth arguing about moves off the multiple and onto the spending. Given the scale of the buildout, are we overbuilding capacity faster than demand can absorb it? His answer so far is that the return on all of it has been positive, with one temporary caveat.
- The "divot" in ROI right now comes from Blackwell, which is an immense expenditure being used largely for training, and training does not generate an immediate return.
- The early agentic checkpoints he points to, GPT-5.2, Grok 4.20, and Codex 5.3, tell him agentic AI has arrived and the eventual ROI on Blackwell will be very high.
He also names the paradox sitting underneath the fear, which is that the market simultaneously worries AI will put everyone out of work and prices the most AI-exposed mega caps at attractive valuations. One explanation he offers is a lingering fear of a global-depression scenario, which he attributes to the analyst Citrini, and which he thinks is unlikely. What he wants, in a genuinely unusual aside, is a strong bear to argue against.
"In that spirit, I'm so grateful to Michael Burry. His substack, it's a godsend, and he's a really smart, intelligent guy who's making a really credible bear case every day. We want that. We want a really smart person credibly banging a bearish drum." Gavin Baker
Wanting a strong bear around is itself a form of discipline. A market with a credible, well-argued bear is a market where the diversity of opinion has not yet broken down, which is exactly the condition that keeps a correct excitement from curdling into a bubble.
Watts, wafers, and scars
The second reason Baker does not expect a 2000 repeat is physical. The economy is short two things that gate the whole buildout, and he keeps them paired.
"We are fundamentally short watts and wafers. And I think that shortage may prevent an overbuild. But even if we solve the shortage of watts with orbital data centers, TSM is still a bottleneck." Gavin Baker
Wafers are the harder constraint, and his description of why is the most investable part of the segment. Chip manufacturing, he says, is like baking. In his account the industry buys largely the same equipment, and where fabs once chose between Nikon and ASML scanners, he says they now lean almost entirely on ASML. The differentiation, he argues, comes from the recipe, the sequence of steps, and the trial and error, so a manufacturer that is ahead is very hard to catch. He credits TSMC's lead partly to Intel making a terrible mistake he attributes to ego, and he does not expect the bottleneck to ease quickly, because the people running the leader are, in his telling, tough operators who once dismissed Sam Altman as a "podcast bro" and will not expand capacity fast enough to close the gap.
The scars are the second lid. Baker draws the distinction between the two crashes in plain magnitudes: 2009 took tech down roughly 60 to 65 percent, while the dot-com bust took it down 80 to 85 percent, and the difference is not a rounding error, it is a different category of pain.
"The scars are so deep from that bubble that just that puts a lid on tech valuations. And I think that's one reason tech has compounded at such a high rate really since Google went public." Gavin Baker
Those two crashes are why he thinks tech has compounded so well since Google and Salesforce went public in the same year, around 2004, because the fear of another bubble kept a permanent lid on valuations and left room to grow into. Underneath the watts-and-wafers point is his first-principles read on why AI is so hungry for both. AI is probabilistic and recomputes its answer every time, even with a harness, chain of thought, or multiple agents, which is what lets it do superhuman things that deterministic software written by humans cannot, and it is also why it is extraordinarily compute-expensive. If AI eats the world, in his phrase, silicon eats the world.
The Tesla first-principles case
Baker's Tesla conviction is old and specific. He missed the IPO, then went on a bus tour of Silicon Valley right before the lockup expired and showed up to a 6 p.m. meeting with Elon at a roughly $1.5 billion market cap that most people skipped. What sold him was a first-principles argument about inputs.
"It's the only car or type of transportation where the core inputs were deflationary. The engine and everything in a car is kind of inflationary over time because we're in a resource-constrained world. But the price of a battery, the energy density of batteries, has been compounding mid to high single digits for a long time, maybe 200 bps below solar photovoltaic cell efficiency." Gavin Baker
The rest of the case follows from the physics of where you put a battery. Because the pack sits on the floor, the car gets a lower center of gravity and a lower polar moment of inertia, so it handles better than an internal-combustion car, and an electric motor controlling traction millisecond by millisecond lets it accelerate faster. Because there is no 800-pound engine block to place ahead of or behind the passengers, you can build front and rear crumple zones, which he argues makes Teslas fundamentally safer, with a much higher survival rate in high-speed collisions over 80 miles an hour. Cheaper over time, faster, better-handling, safer, quieter, more storage, the whole thing made sense to him at once. The method underneath the position is worth as much as the position: he says he has never missed a Tesla or Nvidia public transcript, on the principle that once you identify an exceptional company you stay close to it, which is how he has followed Nvidia for 25 years and Tesla for 15.
The age of Elon
The Elon section is the most openly admiring stretch of the conversation, and Baker frames it as an explanation for a business fact rather than a fan's tribute. The business fact is talent retention. He argues that Elon's companies attract exceptional engineers because the missions are authentic, and that this is an underexplained part of the success.
"If you're a really talented engineer, for a long time your choices were you can go work on making people slightly more likely to click on this blue link for Google search, or click on this ad for Meta, or show this good looking person in Ibiza on Instagram. Or you can focus on decarbonizing the world, which is Tesla." Gavin Baker
The claim behind the admiration is a specific one about acceleration. Baker credits Tesla and Elon with doing more to decarbonize the world than all environmental activists combined, and estimates Elon pulled EVs forward by 20 to 30 years, with US per-capita emissions now at roughly 1925 levels. He also singles out Antonio Gracias as a genuine contributor to the Tesla outcome, pointing to the "Antonio and Tim" chapter in Walter Isaacson's biography, and he ties the whole portfolio of missions together, Starlink, a multiplanetary civilization that can survive an asteroid impact, X, xAI, into what he thinks historians will call the age of Elon. The investable residue is the talent point. A firm that can reliably attract and keep the best engineers on hard physical problems has a moat that does not show up on a multiple, and it is the same quality he credited earlier to Jensen, Hock Tan, and Lisa Su, the rare semiconductor leaders who retain exceptional people.
AI, games, and the compute wall
Because Baker is a gamer, his sharpest read on the near-term limits of AI comes through video games, and it doubles as a caution against straight-lining the technology. He thinks the impact on games has been misunderstood in both directions.
"AI world models are going to lower the cost of developing a game by 90%. If it used to cost $300 million to make a triple-A title like Call of Duty, it goes down 90%. And so this might end up being really bad for companies that make video games, because you're going to have a lot more competition. But if you were a video game platform, this is going to be really good for you because there's going to be an explosion of content." Gavin Baker
The number that anchors the caution is a cost comparison. The idea that games will soon be rendered by AI on the phone instead of by the GPU already in every phone, iPad, and PC, he calls ridiculous for the next five to seven years, and he prices it out: rendering a live game like Monopoly Go using list prices for a model like Veo 3 would cost more than two orders of magnitude above the game's actual revenue. Compute is not free, and the local GPU is not going away on the timeline the most aggressive forecasts assume. That same compute reality frames his coda on human value. The human brain runs on 20 to 30 watts while training a frontier model takes hundreds of megawatts, and in an energy-constrained world he is optimistic that human minds keep their value for a long time, which is part of why he finds Neuralink interesting as an attempt to fix what he calls the human input-output problem. On whether human creativity is truly irreplaceable, he will not overclaim, and he lands on the honest answer he says is correct for almost everything about AI, which is "maybe."
Blowing up on pharma
The most transferable material in the interview is not about AI at all, it is about surviving being wrong, and it comes from the worst stretch of Baker's career. After a fast rise on semiconductors, he was handed Large Cap Pharma at 25, got the sector call wrong, and went from one of the most highly ranked analysts at Fidelity to the bottom. He argues that going through it early was the luckiest thing that could have happened.
"So much of succeeding as an investor in public equities is resilience, tenacity, and loving the game. Forget succeeding, you cannot succeed as an investor if you don't love the game, because you're competing with people who do. You can't survive if you don't love the game, because the love of the game has to be something that keeps you going." Gavin Baker
The reason he frames public equities this way is competitive. Unlike venture or private equity, where you price a deal against maybe ten people, in public markets you compete with everyone on the planet, including a lot of very smart people who are not polished, and that makes it the biggest competitive set there is and a genuinely hard game. What he did with the drawdown became his template. He returned to a touchstone book, "A Wizard of Earthsea," a story about failure caused by ego and the climb back out, and he built a physical outlet on the theory that you need a number you can improve through effort, losing about 30 pounds without trying as the stress reshaped him. In the same stretch he spent an hour or two a day for a year with Fidelity's new team of quants, whom skeptical senior PMs called the "green eyeshade guys," learning quantitative risk management and the common factors that link stocks you would not think were correlated, one of several habits, alongside the deep primary-source work he later did on tobacco litigation, that came out of forcing himself through the failure.
Panic early or double down late
The distilled version of everything Baker learned in that drawdown is a single line from a mentor, and it is the most useful sentence in the conversation for anyone managing money through a downturn. He starts from a hard premise, that being in a drawdown means you are wrong, full stop, quoting the Fidelity PM George Vanderheiden that being early is the same thing as being wrong.
"Ultimately, as an investor, you either have to panic early or double down late. And essentially no one does both. And know thyself. I am not a panic early person, I'm a double down late person. And I think knowing that helps me go through a drawdown or a tough period of performance." Gavin Baker
The value in the line is that it converts a personality trait into a decision rule. He attributes it to a mentor named Jennifer, and pairs it with reminder from a colleague he calls Wymer -- likely Fidelity's Steve Wymer, that it is statistically just as hard to land in the bottom decile as the top, because the game is probabilistic, so a bad stretch is not proof of stupidity. The payoff of surviving several of these, he says, is that decision quality eventually improves during the hard times instead of degrading, because you have a record of having been here before and made good choices, so you arrive at your most confident rather than your most gun-shy. It is the same idea he opened the interview with, that the whole job is finding a philosophy and process that fits your own emotional makeup well enough that you stay rational when the market turns against you.
Be kind, be scrappy, seek hardship
Asked for one piece of advice for a 16-year-old, Baker gives the answer he says he has given at family graduation dinners for years.
"Be kind and be scrappy. A lot of people who are kind are not scrappy, and a lot of people who are scrappy are not kind. Being kind is super powerful. There is karma in the world. If you are kind to people, a lot of people pay it back." Gavin Baker
The reason he frames kindness as strategy is that investing, in his view, is a positive-sum game rather than a zero-sum one. He borrows a lesson from Steve Schwarzman's biography, that you build your reputation to a degree you do not understand in your early and mid-twenties, and that the people in your training class will remember not whether you were the best but how you treated them, which is the network you call on decades later. His own version is a rule about reciprocity, that he will bounce the ball once to anyone and once more if they do not return it, then never again, which sorts you into a group of cooperative, like-minded people, a keiretsu, and he counts Antonio Gracias in his. He closes on a harder note pulled from the science-fiction series Sun Eater, the mantra "seek hardship," because no matter how good you are you will encounter it, and citing the Alpha Architect paper that even God would get fired as an active manager, he argues the decisions you make in those stretches are what define a career.
The bottom line
Baker's answer to the bubble question is really two answers held apart on purpose. There is no valuation bubble, because the multiples say the opposite, tech at a discount to staples. There may or may not be a capex bubble, and the thing that settles it is whether the buildout is funded with cash flows or debt, and whether the shortages of watts and wafers throttle the overbuild before demand can pause. The forward variable he is effectively watching is the ROI divot filling back in as Blackwell shifts from training to inference and the agentic models start paying their way, against the risk that the funding turns fragile or the demand pauses first. What is hard to argue with is the discipline underneath the optimism. A man who wants a credible bear arguing against him every day, and who treats being in a drawdown as definitional proof that he is wrong, is not the profile of someone caught inside a mania. The question he leaves open, and the one worth carrying forward, is whether the watts and the wafers stay scarce long enough to save this cycle from the ending his own history books describe.
Thanks for reading! If you'd like to read more, follow @firesidealpha or subscribe at firesidealpha.substack.com.
