Market Predictions, Rates & Inflation, DOGE, CES, AI Compute | BG2 w/ Bill Gurley & Brad Gerstner

Summary

• The boogeyman is rates, not AI: Gerstner’s team-wide year-end exercise landed on inflation plus higher rates colliding with elevated valuations. The 10-year has gone up almost 100bps to ~4.7% while the market repriced from many cuts to one in the first half of next year; if the 10-year goes to 5.25–5.5%, “it’s going to be an anchor on the stock market.” Altimeter started 2023 at 95% net long, last year ~80% — fewer chips on the table because “the price of entry to play is higher.”
• Not a bubble, but no cushion: S&P at 23x vs. the recent 22x peak and 16x trough; Meta 23x, Google 21x, Nvidia down from 66x to 36x consensus (28–30x on Altimeter’s numbers) while growing at unprecedented rates. The catch: Mag-5 earnings decelerate from 44% to 21% while consensus needs the other 495 to accelerate from 2% to 11% — healthcare +4→+20%, industrials -4→+16%, materials -10→+17% — and tax cuts explain only ~30% of that bump. If non-tech earnings miss, “it’s hard for me to see how you can have a big year.”
• Capex is real and pre-committed: combined large-tech capex stepped from ~$160B to ~$260B, Microsoft alone guiding $80B, consuming roughly 100% of incremental free cash flow at Meta/Microsoft. The debate over whether the 2025 spend materializes “we can put to bed” — it shows up, partly prisoner’s dilemma, but mostly because if Elon is right that AI does every cognitive task within 3–4 years, the labor being replaced “is measured in trillions.”
• Compute stays scarce all year: Microsoft was compute-constrained all 2024; OpenAI couldn’t ship Sora or voice widely for lack of compute; Jensen says inference — already ~half his revenue — goes up “a million x, maybe even a billion times” as reasoning models and agents devour tokens. Brad: “I think we’re going to end this year compute constrained.”
• The tradeable recipients: consensus Nvidia data-center revenue of $61B (2023) → ~$200B (2025) implies share loss Brad doesn’t believe, since the $60B step-up equals the hyperscalers’ incremental spend almost wall-to-wall. Memory is the contrarian leg — SK Hynix at ~6x forward because the market prices commodity boom-bust while Altimeter sees “secular not cyclical” plus a re-rating. Power is the binding constraint: likely Gavin Newsom still hasn’t signed Diablo Canyon past 2029 (10% of California’s clean power) while “China is building 100x as much as we are.”
• DOGE is the swing factor for the 10-year: 2024 federal spending of $6.7T vs. a 2019 baseline grown at 2.5%/yr of $5T — a $1.7T differential that explains the $2T DOGE target. Trump tax cuts add ~$400B/yr of stimulus; the market wants $300–400B of offsetting cuts and “I’ll believe it when I see it.” Brad’s out-of-consensus forecast: this administration actually cuts, the 10-year comes in, and “it could be off to the races” — resolved in a single reconciliation package by April/May.
• State AI regulation is the stupidity risk: the post-1047 regulation push “moved underground” into as many as 25 state initiatives, with the Texas bill (risk assessments, backward-looking liability) the most ironic given Texas’ pro-business brand — Bill calls a state-by-state gauntlet “mind-numbingly stupid” versus federal preemption under the commerce clause. Brad expects Trump to rescind the Biden EO, and 2025 to bring much more open-sourcing across major labs.
• Researchers think they’re “looking into AGI” and are surprised by outside skepticism; reasoning-model scaling is only at a “ChatGPT 2.0 level” on the curve, with dramatic 2025 breakthroughs expected in coding agents from Google and OpenAI. Google faces “the largest innovator’s dilemma in the history of Silicon Valley” but the breadcrumbs (Gemini, deep reasoning, NotebookLM) have Apple worried for the first time; Elon’s xAI is effectively the hyperscaler for a robotics-autonomy-Grok “keiretsu” with a very low cost of capital.

Deep dive

1. The year-end exercise: golden age setup, but fewer chips on the table

• Gerstner’s process, exposed on request: every Altimeter team member independently journals a December-of-next-year view — “you really have to get in the habit of being an analyst, a forecaster, a prognosticator” — then the team reconciles. The bull ledger: lower taxes, lower regulation, strong GDP and employment, plus decade-long megatrends (AI, robotics, self-driving) “that really seem to be bearing fruit.”
• The offset is entry price. Altimeter started 2023 at 95% net long (“all your chips on the table”), started last year ~80%, and is running lighter now because “valuations are quite high relative to where we started ‘23 and ‘24” — invoking Buffett and likely Munger: “the most important thing in investing is the price of entry.”
• Brad’s standing question to the team: “what is the boogeyman — the thing we’re not thinking about?” His answer: smart friends are shorting the US 10-year; inflation plus higher rates on top of high valuations “could put a damper on at least the market in the first part of this year.” Near-term Mag-6 headwinds he thinks are under-appreciated: dollar strength hitting non-USD revenues in Q1, and capex “higher than people think.”

2. Valuations aren’t a bubble — but the earnings baton must pass to non-tech

• The multiple map: S&P peaked ~22x in Q4'21, troughed 16x in ‘22, now 23x — near peak. Meta 23x, Google 21x, Nvidia from a 66x peak to 36x consensus, “closer to 28 or 30 times our estimates” — on unprecedented growth for companies this size. Verdict: “not as good a deal as you got at the start of ‘23 or ‘24, but if we achieve the growth expectations… I don’t think it’s overly onerous.”
• The 2024 anatomy: the Mag 5 (Microsoft, Meta, Nvidia, Amazon, Google) grew earnings 44%; the S&P as a whole 9%; the other 495 companies just 2%. For 2025 consensus has Mag-5 growth halving to 21% against hard comps while non-tech accelerates from 2% to 11% — healthcare +4→+20%, industrials -4→+16%, materials -10→+17% — with tax cuts explaining only ~30% of the bump.
• The conditional call: to get 10–15% on the market you need that non-tech acceleration to be delivered and rates to stay below 5%. “If those earnings aren’t delivered… it’s hard for me to see how you can have a big year.”

3. The capex supercycle: prisoner’s dilemma or trillions of labor value

• Combined capex at Google, Meta, Amazon, Microsoft, Apple, Oracle stepped from ~$160B to ~$260B in 2024, consuming “the vast majority of the incremental free cash flow” — Meta and Microsoft above 25% of revenue and bumping against 100% of FCF, Apple the outlier falling toward 5%. Microsoft’s $80B number caused “a bit of a kerfuffle” but is basically his existing run rate; even Elon called it “a big number in anyone’s universe.”
• Brad’s taxonomy: it’s either an offensive bet on future revenue or a defensive one — “I can’t not do this because my competitors are doing it” — and the wall of worry over returns is precisely why these multiples haven’t gone higher. Omaha “is pretty skeptical,” seeing a telecom-buildout rerun. Bill’s reflexivity question got a rebuttal: Brad called the optics theory “fairly conspiratorial” and pointed to compute constraints — if Satya said $60B instead of $80B, would the market read it as losing conviction or share?
• The justifying math comes from Elon at CES: “within the next few years we’ll be able to do any cognitive task… I’d say max three or four years.” Brad: “if you believe that to be true, the value of all that human labor that you’re replacing is measured in trillions” — which is why the CEOs have decided they can’t not spend, “even though it makes the CFOs’ voices a little high-pitched.”
• On why Wall Street can’t model this: in 2023, 26 analysts covered Nvidia and “the consensus forecast of all 26 of them missed by 80%… that’s how wrong you can be if you’re thinking linearly at a moment of a phase shift” — the same early-cloud window (2012–13, not 2016) that made Snowflake and Okta possible.

4. Compute constraint doesn’t end in 2025

• The demand evidence: 300M+ weekly ChatGPT users, Microsoft compute-constrained all 2024 with inference revenue set to accelerate in H1'25 as constraints ease, and Sam saying publicly OpenAI couldn’t widely release Sora or voice — Pro pricing exists partly “to artificially reduce the demand because they don’t have the compute.”
• The kicker is reasoning: O-series models and Google’s deep reasoning are “compute hogs” — token generation, prompt repopulation, branching inference “dwarfs single-shot ChatGPT,” before adding personalization, long-term memory, and actions. Jensen on the pod: inference, already ~50% of his revenue, goes up “a million x or a billion x.” Brad’s categorical call: “I don’t even think 2025 gets us to compute surplus — we’re going to end this year compute constrained, and a lot of the frontier labs feel the same.”
• Bill’s business-model observation: Grok 2 stays free but Grok 3 may be paid, Google launched a paid high-end tier — if the frontier converges on subscription pricing rather than free-like-search, “that could start to get sticky” for OpenAI (rumored $4–5B run rate exiting the year, plausibly $10B+ ahead). Brad’s cut: Mistral has “already raised the white flag,” and Anthropic at reported sub-$1B, mostly enterprise API, faces the question of whether it can afford to keep up — “you cannot go out and spend and not have a business model to support the spend.”

5. Following the pre-committed dollars: Nvidia, memory, power

• The Nvidia arithmetic: consensus data-center revenue goes from $61B (2023) to ~$200B (2025); the $60B increase from ‘24 to ‘25 equals the hyperscalers’ incremental spend, and since custom ASICs are “tiny on a relative basis,” consensus arguably implies Nvidia loses share — “I don’t think that to be true.” The stock has been flat since June amid pre-training-wall debates and DeepSeek’s small-model scare; Brad’s discipline: “every time you have to recalibrate based upon the facts on the field.”
• The contrarian leg is memory: SK Hynix (with Gavin Baker) and Micron, on a high-bandwidth memory shortage “for as far as we can see” in an inference-time-compute world. SK trades ~6x forward because the market prices commodity boom-bust; Altimeter’s view is “secular not cyclical,” software-laden product — two ways to win: earnings growth plus a non-commodity re-rating.
• Power is the primitive: operators tell Brad “I know this is hard to believe, but we are power limited” — capex they can’t bring online for lack of megawatts, while “China is building 100x as much as we are.” His concrete ask, pushed with PG&E’s COO and likely David Sacks: likely Gavin Newsom still hasn’t signed the Diablo Canyon extension beyond 2029 — 10% of California’s clean power — “it’s insane… at a minimum” it must be extended this year.
• Why Altimeter doesn’t chase energy stocks despite peers’ returns: the North Star is essentialism. “I would ask my analysts, why is that better than Nvidia? It’s a derivative Nvidia — the exact same bet, so why not just buy more Nvidia?” Bill’s coda: “there’s a fool in every market, and if you don’t know who it is, it might be you.”

6. Rates: the market is calling Congress’s bluff

• The puzzle: headline inflation went from a nine-handle to a two-handle (Morgan Stanley consensus 2.2% by year-end), yet the 10-year is up ~100bps to ~4.7% and the market has repriced to one cut in the first half of next year. Brad’s read: ~$400B/yr of stimulus if Trump tax cuts pass, plus regulatory relief, risks reigniting inflation — and crucially, “the market is saying we do not expect Congress to have the courage to cut an equivalent amount out of the deficit… I’ll believe it when I see it.”
• The binary he lays out: if hundreds of billions get cut, “the market will say great fiscal discipline,” the 10-year comes in, “and it could be off to the races.” If not, and the 10-year goes to 5.25–5.5%, “it’s going to be an anchor on the stock market, an anchor on the economy” — with the bond vigilantes as enforcers. The whole question gets resolved in one reconciliation package by April or May.

7. The DOGE math: $1.7 trillion above baseline

• Altimeter’s own analysis: 2019 federal spending was $4.4T; grow every category at 2.5% (GDP/inflation/population proxy) and 2024 should have been ~$5T — actual was $6.7T, a $1.7T differential. “When you hear DOGE talk about $2 trillion in savings, this is really what has people optimistic.” A bipartisan proposal already identifies $700B of “easy deficit reduction” over 10 years — a start, not the answer. Trump could even act unilaterally by challenging the Impoundment Control Act.
• Bill’s pushback — worth keeping: healthcare per-capita spend has grown well above the 2.5% baseline for structural reasons, so you can’t just decree less spending — you have to attack the industry problem. Brad concedes the global demographic squeeze (fewer workers per retiree) means either delivering less than promised or more efficient delivery via technology and AI, plus smarter defense spend — “the track record of Congress, both Republicans and Democrats, has not been good.”
• Bill’s meta-point on why DOGE has room: “I don’t think there’s anyone anywhere that would stand up and say our government is great at execution” — Elon calls it “shooting fish in a barrel” — and the bigger unlock may be repealing restrictive policy to lift GDP growth itself. Brad: getting to a balanced budget by 2029 at ~$6T “is just not that hard,” and versus Europe — “a basket case and a disaster” — the US just needs “the courage of our conviction.”

8. State-by-state AI regulation is the own-goal to watch

• The post-SB1047 regulation push “moved underground” into as many as 25 state-by-state initiatives, with Texas hottest — bitterly ironic given Abbott and Elon made it the least-red-tape state. Bill on the substance (citing Dean Ball’s analysis): mandatory published risk assessments plus backward-looking liability — “did I run the analysis where I should have known about the risk?” — applied broadly. His verdict on a state patchwork while Washington frets about China: “mind-numbingly stupid,” said with the caveat that legislators might quote him.
• Brad’s counter-frame to the “Silicon Valley wants no regulation” boogeyman: the pushback is to uninformed regulation — “if you have this patchwork where states are onerously regulating these companies out of the gates, all you’re doing is handing the gold medal to China.” Both want federal preemption under the interstate commerce clause — “one rule of the road” — and Brad expects Trump to rescind all or most of the Biden AI EO, with Sacks as the coordinating function. Brad’s adjacent 2025 prediction: far more open-sourcing across major players (Microsoft’s open-source models last week “really surprising”), because there’s “a lot more commonality between the major players around AI than there is division.”

9. Google’s innovator’s dilemma vs. the perfect AI basket; xAI as keiretsu

• Bill’s Google case: Gmail, Docs, Android, and Slack-/Zoom-competitive products — “the perfect AI basket of assets” nobody else has, evidenced by Jason switching to a Pixel. Brad’s sizing of the other side: “the largest innovator’s dilemma in the history of Silicon Valley… it’s almost impossible to replace a 99% incremental-margin monopoly search business” because whatever comes next won’t be a monopoly or carry those margins. But since late November the breadcrumbs — Gemini, deep reasoning, NotebookLM, resilient search volumes at travel/lead-gen customers — show Sundar “starting to get feisty,” and Brad suspects Apple is worried about Google on the phone “for the first time” — a Gemini assistant that books your hotel could be a 10x better product. Bill’s practical tell: Gemini’s local-review corpus gives it an advantage (“what three dishes do people love, which two to avoid”).
• On xAI, Brad refuses Bill’s framing that Grok 3 is a pass/fail test of pre-training scaling (Bill holds his ground: if it beats OpenAI’s latest, “that would be a new data point”). Brad’s structure argument: Elon runs three AI vectors — robotics (“millions of humanoids by 2028”), autonomy (his new Tesla “drove me home from San Francisco last night”), and Grok — across a Japanese-style keiretsu of companies that share insights, so giant coherent clusters aren’t wasted even if pre-training plateaus: they convert to reasoning training, post-training, inference. Plus Elon’s advantages: standing up compute on “crazy timelines” and a very low cost of capital. “It’s so great for the US ecosystem that Elon is keeping Anthropic and Google and OpenAI honest.”

10. Coopetition, Bell Labs, and staring into AGI

• Bill’s historical observation: in the Wintel era “people stayed in their lane”; now Microsoft touts its own models while partnering with OpenAI, hyperscalers build Trainium/TPU while buying Nvidia, and Nvidia ships models and AI PCs against its own customers. Brad’s explanation: frontier capital needs were “nuclear level,” beyond VC, forcing labs into hyperscaler embraces — Google hedged with Anthropic, Amazon bet Anthropic, Microsoft OpenAI — and the also-rans (Inflection, Character) could get subsumed. He’d underweight the drama: Nvidia’s models aren’t a frontier play, and his macro takeaway is nearly $300B of non-governmental R&D — a Bell Labs moment — “so damn bullish” for Team America, from companies that were recently criticized for buybacks.
• The closing signal on reasoning: lab researchers “feel like they’re looking into AGI and they’re surprised how skeptical the world is” — the dissonance comes from seeing unreleased, compute-starved capabilities. On the scaling curve, reasoning models sit at roughly “a ChatGPT 2.0 level,” implying big 2025–26 gains; Brad expects real coding agents from Google and OpenAI, meaningfully better than today’s prompt-injection-based startups, then consumer memory-and-actions (an Altimeter partner’s booking.com demo on Anthropic computer use was “still fairly embryonic — now build that on the back of o3”), with pre-trained and O-series blending into one model. Bill’s honest hedge to end: he’s not a macro fan — “it falls in the too-hard bucket” — and an open invitation to listeners doing interesting things with chain-of-thought to write in.