Adam Parker on the AI Supercycle, NVIDIA and Where Investors Still Have an Edge

Summary

• Parker’s headline call: Nvidia’s earnings should at least double by the end of this cycle, with a $10 trillion market cap “an easy thing to see” versus roughly $5.5T today. The moat is CUDA — a software platform that lost money for 15 years and is now an installed base where “you can’t really do any complex inference problems without CUDA” — and 15% annual growth for five years, a “very conservative” revenue view, gets the stock roughly 70–80% higher even with continued multiple contraction.
• On cycle timing, his conditional framing is that we’re only “three years four months into a 10-year investment cycle” dated from Nvidia’s first big upward sales revision in May 2023 — “why the hell would the stock already have peaked?” His proof that AI is younger than it feels: a summer associate found the letters “AI” appeared in none of the eight 2023 year-ahead outlooks from Goldman, Morgan Stanley, J.P. Morgan and UBS.
• He thinks the sell-side’s earnings estimates are now biased low — a reversal of the pattern, dating from 1978, in which January estimates were too high in 80–85% of years — with 2028 the most underestimated year. Only ~20% of US equities even mention cost-related AI productivity today; his S&P earnings math could move from roughly 455 to 475–500, and at 19–20x “I don’t see why you’re not at 10,000 on the S&P.”
• Valuation doesn’t pick stocks; getting the E right does, and the multiple now carries information. Cheapest-decile stocks (~8x forward) that just got cheaper beat estimates only 48% of the time versus 75% for expensive stocks getting more expensive, and “the cheap stocks missing have never been punished more.” Michael cites Micron as the E-error exhibit: two years ago analysts modeled something like $6B of free cash flow over eight quarters; they now think it is around $300B. Parker adds that valuation is also often assessed independently of the balance sheet, which it should not be.
• The unpriced bull case is physical AI made in America: humanoids plus the silicon and software inside them driving “a kind of sneaky big re-industrialization,” and healthcare possibly the best-performing sector of the next decade. “The market’s telling me there’s 0% chance and I think it’s 30 or 40… I want to arb that difference.” His three five-to-ten-year above-GDP convictions: healthcare services, compute, and power — while semis, ~2% above GDP for 30 years, “might be eight or nine or 10% above GDP for 10-15 years.”
• The meta-lesson of the episode is edge awareness: “if AI hasn’t made you a little bit worried about your skill, you might be a psycho.” Don’t compete with a hundred pod-shop analysts leveraged 6–12x and turning books 600–800% a year on the quarter — “everyone’s playing a different game,” and a long-duration diversified book is a different game where an investor may have an edge.
• A tradeable structural warning: ~700 levered/inverse ETFs exist today versus 45 three years ago, ~60 have already failed this year, and about half of all options are retail zero-day. Triple-levered semis (SOXL) has delivered roughly 220%, not 3x, with Parker saying its asymmetry may be closer to triple-levered on the downside than the upside — and “if we get a more austere year or two, some of that will get cleared out.”

Deep dive

1. Flexibility first — and knowing where you have no skill

• Parker’s opener sets the tone: “in the last year, I feel like I’ve written 15 times about how there are things that I now believe that I used to believe the opposite” — flexibility and adaptability matter “more than any time in my life.” The self-audit that follows: “if AI hasn’t made you a little bit worried about your skill, you might be a psycho.”
• On the multistrats: if a hundred analysts at every pod shop are gaming one stock’s quarter, leveraged six to 12 times, “you’re not competing in a game where you have an edge.” Michael’s version: a long-only investor can hold Palantir on Alex Karp’s vision — Rule of 40 running 127, 144, 145, 155 — without calling the quarter, a luxury the pod seat doesn’t have.
• Parker’s darker aside: “one thing I feel like I know, without any arrogance, is more than a lot of people” — people who made a ton of money for themselves while they didn’t make any money for their investors. Gun it for $17M one year, blow up the next, move to the next multistrat. “The arbitrageable thing is that everyone’s playing a different game.”

2. Valuation doesn’t pick stocks — the E does

• The regime shift: unconditional probability of beating estimates is ~72–73%, but “the penalty for missing estimates for US equity right now is really harsh,” far exceeding the reward for beating. The old play of buying a business with temporarily suppressed margins and multiples 18–24 months early is now much harder to justify: Parker would rather not own a name he knows will miss. “The cheap stocks missing have never been punished more.”
• His decile stats: cheapest-decile stocks (~8x forward) that just got cheaper have about a 48% chance of beating; names around 30–33x forward earnings that just got more expensive have a 75% chance. “Now there’s information in the multiple… that didn’t used to be” — and nothing in the sub-10x bucket is cheap for mysterious reasons.
• Michael’s Micron example: two years ago, the sell-side thought it would generate something like $6B of free cash flow over eight quarters; it now thinks the figure is around $300B. Parker’s broader point is that investors often assess valuation independently of the balance sheet, even though Micron’s cash position and expected free cash flow should affect the valuation.

3. Semis: from cash incinerators to the market’s most important sector

• The origin story: Parker launched Bernstein semiconductor coverage on October 1, 2002, after 11 months in a room researching seven stocks, under the title “share gainers and margin expanders are multiple expanders” — a thesis that, unlike much of his output, “aged okay.” Nvidia was too small by market cap and liquidity to cover; 25 years ago respected buy-siders would have put Jensen in the “dishonest and incompetent quadrant.”
• The Micron parable, told in full: over a five-year stretch it burned “exactly $365 million per year” of free cash flow, and his sarcastic mid-2000s note argued handing a million dollars a day to the employees in the parking lot would have beaten making DRAM in Idaho. Micron had 9,000 employees in Boise, where roughly 200,000 people lived. Consolidation plus the AI supercycle later, it’s a trillion-dollar market-cap company — “incompetent, dishonest and terrible businesses can turn into the most important securities in the world.”
• To Michael’s deep-cyclical challenge: most semis aren’t deep cyclicals anymore because most no longer manufacture — Nvidia has Taiwan Semiconductor make its chips — though memory and DRAM stay more cyclical. “The periodicity and the amplitude are way different than history.”
• The macro frame: “this is a supercycle. Compute is going to grow way above GDP for a really long time.” Semis ran ~2% above GDP for 30 years; it “might be eight or nine or 10% above GDP for 10-15 years” at higher margins for US companies — while we’re short compute and short power for the compute.

4. Nvidia: three years into a ten-year cycle

• The misremembering setup: his summer associate read the 2023 year-ahead outlooks from Goldman, Morgan Stanley, J.P. Morgan and UBS — “the letters AI were not in any of those eight documents.” If May 2023 was Nvidia’s first big upward sales revision, “you’re three years four months into a 10-year investment cycle. So part of me wants to say, well, why the hell would the stock already have peaked?”
• CUDA lost money for 15 years and is now an installed base where “you can’t really do any complex inference problems without CUDA” — a lead that will be hard to replace.
• He acknowledges the circular-lending critique — “if you buy a hotel and then you sleep in it and count your own bed as occupancy” — but: “I don’t see how their earnings aren’t at least twice as high at the end of this cycle.” 15% a year for five years is a double on a “very conservative” revenue view. He expects substantial cash accumulation without a huge free-cash-flow-to-net-income disconnect over the cycle; $10 trillion “would be sort of an easy thing to see at the end of the cycle” versus ~$5.5T now — call it 70–80% higher, embedding continued multiple contraction.

5. The estimate skew is to the upside — and S&P 10,000

• Michael’s “Sand Hill Road” question — what if you’re wrong to the upside? — is the one Parker says most people don’t ask. Since 1978, analysts’ January estimates were too high in 80–85% of years (14% forecast, ~8% delivered); since Nvidia’s 2023 revision, “estimates have been consistently too low.”
• 2028 is where he thinks consensus is most wrong: “a lot of people are just embedding this decline” precisely when AI productivity ramps — his natural-language processing work on transcripts shows only ~20% of US equities even cite cost-related AI productivity today, and “there’s going to be way more than that by 2028.”
• The math: 18% earnings growth next year plus another 10% in his calculation gets roughly 455 in S&P earnings; “I think there’s a real chance it’s more like 475 or 500,” and at 19–20x, “I don’t see why you’re not at 10,000 on the S&P.” On funding fears, he quotes a smart investor on Musk: “a superpower which is unlimited access to cheap capital” — though “I don’t want to whistle by the graveyard.”

6. What could go wrong: white-collar jobs — and what replaces them

• Michael’s risk scenario is that three, five or six years out, white-collar jobs in the $100–500k range could see “10-15% unemployment.” He tells Parker he does not want to claim unemployment will not be a problem in five years because that “could age like milk.” Michael also adds the week’s news: three major investment banks told the top-100 law firms to restructure pricing because LLMs are taking over work billed by paralegals, first-year associates and junior partners.
• Parker sees law going bimodal — niche boutiques and giants paying rainmakers “like NBA players now, $30 million bucks a year,” with 1,000-attorney middle firms hollowed out, consolidating, perhaps involving private equity or public stocks. His barb: lawyers still have “a lot more differentiated skill than investment bankers on average” — Claude can build a pro-forma P&L, so “the value-add of some whippersnapper from Wharton is kind of low.”
• On education ROI: for the median person college is now “a dumb idea,” great only at the 80th–90th percentile — “your average HVAC employee makes more than your average Harvard grad” — and physical skills stay in demand.

7. Physical AI, humanoids and the unpriced bull case for America

• The geriatrician example, as told: we’re short people to care for an aging population, so “would you let a humanoid take care of one of your parents in their 80s? I think the answer is yes” — sensors measuring vitals, tracking medication, summoning a driverless vehicle, no theft accusations and no drug-delivery mistakes. Michael’s timeline: early adopters buy one within three years, everyone within ten; Parker: “yeah, maybe… I think you’ll have a lot of them.”
• The macro payoff: “the bull case for America that is nowhere near in prices is we’re going to make that stuff in the US” — the humanoids, the silicon, the software — “a kind of sneaky big re-industrialization of America,” with metals, copper, power, sensors and maintenance jobs filling the void left by “paper-pushing jobs.” “I don’t want to get dreamy, but I would assign a higher probability to that than the consensus.”
• His three five-to-ten-year above-GDP convictions: healthcare services, compute, and power — and the arb: “there’s a real chance that healthcare is the best performing sector in the stock market over the next 10 years, and I think the market’s telling me there’s 0% chance and I think it’s 30 or 40… I want to arb that difference.” Go click Nvidia’s healthcare tab — “Jensen knows a lot about healthcare” — and low-margin, employee-heavy businesses, with Quest and Labcorp as examples, could become far more efficient.

8. Bearish sounds smarter — and the data-center narrative has to change

• On the perennial US-debt question he’s fielded 50 times a year for 15 years: “you always sound smarter when you’re bearish.” The only real question is whether it should change positioning today — “the answer almost every day for the last 15 years was no,” and “you get fired if you position your fund for that.” Michael’s gentler version: no data, no path, outside the investing horizon — or just don’t own equities.
• Parker’s water comparison: water use at a mid-size data center and a golf course is close; the biggest data center in the world “would be something like three golf courses” — yet politicians of both parties “act like you’re using the entire Pacific Ocean.” He says 23 states now view a data center as something they do not want in their state.
• His Bloomberg-surveillance answer: “the narrative simply has to change… Sam and Dario kind of scared everyone, and luckily Jensen and Mark are going back and cleaning it up” — people love Netflix, same-day Amazon delivery and Starlink; “that’s what a data center is.” His favorite metric, from asking Mayor Bloomberg how to measure a good mayor: citizens’ “productive years” — “the whole point of all the AI stuff is that we live longer, we’re more productive while we’re alive.”

9. What real portfolio strategy is — the career arc and the retail newsletter

• The speedrun, with Michael noting Parker’s earlier No. 1 ranking: Bernstein semiconductor analyst → Morgan Stanley chief US equity strategist (44 conferences, 32 travel weeks a year — “Bernstein going from MIT football to Alabama football”) → Eminence Capital under Ricky Sandler doing risk and position sizing → Trivariate Research, whose institutional business launched in May 2021 and now does custom risk work on roughly 60 client portfolios. Running his own shop means washing the coffee mugs — and learning “who has pricing power over you and who doesn’t” when UnitedHealthcare emails a 9% hike with “no negotiation.”
• His definition of strategy is not index targets — “I know I don’t have any skill making a one-month stock market call. I also know that nobody else does, but they act like they do” — but stop-loss rules, name counts, adding to losers versus winners. Today’s example: compute an AI beta and a non-AI beta per stock, because “we have the most negative beta stocks in the history of the stock market” — a 10% Exxon position screens negative-beta and “the metric is meaningless.”
• The retail arm — Trivariate Research, at ~$110/month or $1,200/year — grades ETFs “Dave Portnoy pizza style” on whether they deliver the exposure claimed: MTUM “underperformed the S&P while momentum was awesome,” hence the poor grade.

10. Levered ETFs and zero-day options are why the tape is so violent

• Michael’s ETF taxonomy: long-term compounders (VOO/SPY-type), structurally inefficient products (buffers, covered calls — trading compounding for near-term certainty), and the “uninvestable” doubles and levers. Parker’s data: ~700 levered/inverse ETFs today versus 45 three years ago, and ~60 have already failed this year — “more levered ETFs have closed this year than existed cumulatively up until three years ago.”
• The mechanics: ~65% of US equity options are zero-day and ~70% are retail, so roughly half of all options are “retail dudes doing zero-day options,” many on triple-levered products; with roughly 80% of S&P volume at the open and close, big intraday moves are hard to true up. SOXL has delivered roughly 220%, not 3x, and Parker thinks its asymmetry may be closer to triple-levered on the downside than on the upside.
• The behavioral kicker, via Michael’s dad’s 20x Amazon-because-he-loved-the-Kindle position: people “go back to what made the money” — but “if we get a more austere year or two, some of that will get cleared out, I think.”