DeepSeek, Open Source, Tariffs, DOGE, Market Impact | BG2 w/ Bill Gurley & Brad Gerstner

Summary

• DeepSeek’s moment of maximum parity is already behind it. Gerstner’s team, speaking directly with DeepSeek, confirmed the ~$1B total TCO alongside the real $6M final training run (vs ~$10–15M for o1 — “30 to 50% more efficient”), but the log-linear scaling math now demands ~10x the compute to reach o3-class — with no Blackwell access, export controls binding, and most of their limited cluster consumed by inference demand. “This was the moment in time where their compute comparison to OpenAI was the closest it’s ever going to be.”
• Open source may have hit its tipping point. Hugging Face went from 500 R1 variants 48 hours post-release to 1,300 today under a no-strings MIT license, and Sam Altman conceded in his AMA: “I personally think we have been on the wrong side of history here.” Gerstner predicts closed labs open-source one-to-two-generation-old models this year while true frontier models — from OpenAI and DeepSeek — stay unreleased, with “only the agent” shipped.
• Gurley’s contrarian core: a foreign, open-source disruptor beats a domestic proprietary one — for safety, free speech, and innovation — and China is a 30-year open-source veteran (Linux donors, RISC-V) precisely because “if you believe you have the fastest, cheapest, most capable engineers… you’d rather live in a world where there’s no IP protection.” His warning: “Linux doesn’t have a geography” — the winning model may be a fork nobody’s country owns, and walling off the US could hand the rest of the world (DeepSeek at ~15% of ChatGPT DAUs in India) to China.
• Don’t be lazy in believing the “Bessent consensus” that tariffs are just a loaded gun rarely discharged. Gerstner’s variant view: Trump holds a decade-old, principled McKinley-era belief that pre-1910 tariff revenue built America and that swapping tariffs for income tax “gutted the middle class” — meaning tariffs as permanent tax-reform architecture, not negotiating theater. Gurley’s rebuttal: he doesn’t know of a long-term successful tariff program, the Fed’s 2018 washing-machine study showed prices up and few jobs created, and “you’re just going to make yourself Europe.”
• DOGE runs on two tracks, and track two could reach the Supreme Court. Track one is normal reconciliation (a signable bill by April/May); track two is Elon cutting by executive authority — asking “who sends out the wires,” shutting USAID (“no apple to be saved, it’s a total ball of worms,” $50B/yr). Gerstner expects a Schumer lawsuit on Congress’s power of the purse but thinks it’s “on pretty weak footing” against the doctrine he called “empowerment” — and if a credible $1T cut back to the 2019 baseline lands, “interest rates are going to come down.”
• The trade: fundamentals on fire, multiples at risk. Stacked technological, economic, and political uncertainty forces the discount rate up and multiples down even as Jevons-style demand explodes — “I think the super cycle is on fire… but at the moment it’s the Golden Age of uncertainty.” Gerstner tells his team to downsize risk (“you don’t do the tenth best idea”), expects heavy volatility for six months, and warns mid/late-stage VC marks lag but follow public multiples downhill.
• Second-order trap for NVDA bulls: loudly claiming DeepSeek secretly had more GPUs invites Washington to tighten sanctions that hit Nvidia’s own revenue — “you think you’re protecting Nvidia… and you may end up with sanctions that end up hurting Nvidia’s revenues.” A threatened Taiwan chip tariff is just a tax on Nvidia, AMD, Meta and Amazon that buys less AI research — unless structured as a 2.5-year-delayed lever tied to US fab buildout.

Deep dive

1. R1 was real innovation, not a copy — and the founder is the tell

• Gurley’s week-later verdict: DeepSeek genuinely innovated — putting more experts simultaneously against a problem by separating parameters into smaller counts, yielding something cheaper and faster “which no one else had done before” — and it’s “important to recognize that they did innovate versus just copy, because there’s a lot of noise around this whole thing.”
• Two under-appreciated details: they went around CUDA — which Gurley reads as classic bare-metal optimization, like AWS ripping layers out of Linux in the SaaS transition — and the founder interview (likely Liang Wenfeng, translated from Chinese) makes it “impossible to read that and say this isn’t an exceptional founder.”
• Gerstner’s team knows the DeepSeek team directly and confirms the frame: a decade on the problem, hedge-fund wealth applied to deep learning — “it reminds me a little bit of the Jim Simons story at Renaissance.” OpenAI’s only surprise, when Gerstner asked: that a Chinese company got there “before Meta and others.”

2. The compute wall: DeepSeek will never be this close again

• The numbers, confirmed from the company: total compute cost including power lands near the $1B TCO figure, comparable to peer models; the $6M final training run is also correct, versus roughly $10–15M for o1 — “apples to apples they were 30 to 50% more efficient,” from real algorithmic breakthroughs, though achieved six-to-eight months later. Gurley adds the Dylan/Nathan claim from the Lex pod that DeepSeek’s API prices at ~1/12th of OpenAI — with the caveat they may be subsidizing: “Lyft endured rides well below cost for a decade.”
• The core structural call: their o1-era compute stack (~30–50k previously purchased GPUs) wasn’t much smaller than OpenAI’s, but log-linear scaling means 10x the compute for the o3 step — with no Blackwell (a 2–3x improvement on Hopper), export controls binding, and most of their constrained cluster now eaten by inference demand from their own success. “This was the moment in time where their compute comparison to OpenAI was the closest it’s ever going to be.”
• Gerstner’s falsification test: if DeepSeek somehow catches o3 and deep research without GPU access via architectural breakthrough, “that will to me be the definitive statement that they have broken the paradigm.” Gurley’s counter-current: constraints breed innovation (“the reason Llama probably didn’t do it is they had access to scaling”), the paper is public, and “I’m 100% certain people at Anthropic and OpenAI are studying what DeepSeek did” — the borrowing is bidirectional, and the likely Jevons’ Paradox means cheaper just means more.

3. The open-source case: 1,300 forks and Linus’s law

• Gurley’s tweeted thesis — a better world if the most disruptive LLM is foreign and open source rather than domestic and proprietary — rests on transparency yielding “more understanding, more safety, more security, more free speech.” A singular proprietary model concentrates control risk; recall OpenAI itself was founded, per Elon on stage at launch, “to defend against singular control by Google of a closed tyrannical AI.”
• The proliferation data: 48 hours after R1 posted, Hugging Face had 500 variants; 1,300 by this morning — Linus’s law in action (“given enough eyes all bugs are shallow”), with “the R&D force becomes the world.” Enterprises (Levie, Benioff) cheered, it was quickly deployed in AWS and Azure—Microsoft’s deployment was “shocking” given the OpenAI contract dynamics—and it helps everyone down-stack too, from TPU challengers to inference-layer players like Fireworks, in which Gurley says they’re an investor.

4. Why Llama lost the moment — and Altman’s concession

• Gerstner’s pointed question: a year ago open source meant Zuck and Llama; why is the evangelism now about R1? Gurley, an original Red Hat investor (~1999): every commercial open-source player runs “this weird game” of wanting proliferation plus a proprietary clawback — Meta’s “if you use too much you got to come see us” clause (Amazon pays for Llama use). DeepSeek simply chose the MIT license, “pretty against the rail” — the most open available today.
• Altman’s AMA response — “deep seek’s an impressive model… I personally think we have been on the wrong side of history here and we need to figure out a different open-source strategy” — validates R1 in Gurley’s eyes: “whenever a threat is made to OpenAI, Sam tends to go towards it.” So did co-investors rushing to X to cry cheating or demand government action: “you wouldn’t take the trouble if this thing wasn’t real.”
• Gerstner’s structural prediction: closed labs open-source trailing models (he can see OpenAI open-sourcing o1 to fight R1 head-to-head) while today’s open players — DeepSeek included, and Zuck has reserved this right — withhold the true frontier, releasing “only the agent.”

5. China has been playing the open-source long game for 30 years

• Why is China pro-open-source? “For the past 30 years the West has done nothing but accuse them of being IP thieves” — if your engineers are the fastest and cheapest, no-IP-protection is the world you want. Check the Linux Foundation donor list; and “every time we put more constraints on what they can get to, they invest more in RISC-V.”
• Gurley’s history lesson: in 1820 an open China held 33% of global GDP; by the end of Mao’s inward-turned reign, 5%. “If you pull the wall up and we don’t support open source and they do — and everybody else kind of likes them leading that way — that could be a dangerous situation.” Fareed Zakaria’s two takeaways he endorses: the “US two years ahead” line was “hubris” (it’s six months, maybe three, and closing), and open-vs-closed can hit a tipping point where “water runs downhill.”

6. Silicon Valley’s biggest divide: China hawks vs. constructivists

• Gerstner maps the camps: hawks like Alex Karp treating this as “an existential holy war,” versus constructivists — Gurley, Gerstner himself, “frankly Elon” — who think slowing China is a losing battle and the focus belongs on speeding America up: power generation, deregulation. Gurley subdivides the hawks into three: the Dario camp begging for lockdown, the VC-backed defense companies — “I call them the new neocons,” whose tension with China “actually increases revenue” — and people simply “raised to be anti-China.”
• Gurley’s additions to the risk ledger: protection causes harm — “Detroit is not globally competitive anymore, and putting tariffs on BYD’s cars is not going to make Detroit more competitive”; the rest of the world may be “perfectly fine buying $10,000 BYD cars and using DeepSeek”; Congress already has proposals “that would kill open source, that would say we couldn’t use variants of R1”; and the darkest tail: “if you just poke China enough… you increase the odds that they make a run at Taiwan.”
• Gerstner’s clarification of Gurley’s provocation — “I want Team America to win” — met by Gurley’s deeper point: “Linux doesn’t have a geography.” With 1,300 MIT-licensed forks, the winning model may have no sovereignty at all, which doesn’t satisfy the America-wins goal either.

7. Tariffs: the Bessent consensus vs. Gerstner’s variant view

• The week’s tape: 25% tariffs on Mexico/Canada and 10–15% on China were set to take effect over the weekend; Kevin Hassett reframed it Monday as “not a trade war, a drug war” over fentanyl; both countries committed 10,000 troops to the border, tariffs delayed 30 days, and the market now prices them never hitting. The consensus frame is Scott Bessent’s own letter language: Trump’s tariff strategy is “a fully loaded gun, but rarely discharged.”
• Gerstner’s warning — don’t be lazy: Trump’s decade of speeches reveals a principled belief in tariffs. He thinks McKinley was among the best presidents, that America peaked during peak tariffs circa 1880, and that the post-1910 swap of tariff revenue for income and corporate taxes “gutted the middle class.” The back door was explicitly left open: “we may revisit as part of a tax reform strategy.” That’s not a negotiating stick — that’s architecture, a “Trump New Normal.”
• Gurley’s caveat on reading any of it: even if tariffs are purely a negotiating tool, Trump “can never say that out loud or it would take away their ability to be used in that way” — so obscurity is structural.

8. Gurley’s globalist rebuttal: washing machines, wages, and a chip-tariff trap

• The evidence base: Gurley researched it and “I don’t know of a long-term successful high-tariff program” — even McKinley’s success stories ran ~four years. The Fed’s 2018 washing-machine study: prices rose on tariffed washers, on untariffed dryers sold alongside them, and even domestic producers raised prices into the new umbrella — with very few jobs created.
• His blunt wage argument: postwar America scaled because Europe and Japan were decimated and Americans would climb the ladder; today workers in China, Vietnam, Indonesia and Mexico will work 996 “for a wage that is radically lower,” moving up from below the average American — “I don’t know how as a humanist you can say they don’t deserve that.” Gerstner’s pushback: a president’s oath runs to Americans’ living standards, not the globe’s. Gerstner’s point: even with that lens, “trying to get people to make a $40 microwave in America is going to fail… you’re just going to make yourself Europe.”
• The tradeable specific: a threatened tariff on Taiwan is a tariff on chips America cannot yet make (no domestic 2nm/3nm fabs) — a tax on Nvidia and AMD and on Meta and Amazon as buyers, meaning fewer chips and less AI research, “self-defeating” if winning AI is goal one. Gerstner’s better design: a 50% chip tariff delayed 2.5 years with fab-buildout hurdles — a negotiating lever, not a permanent tax — reducing Taiwan dependency.

9. DOGE: two tracks, one constitutional collision

• Gerstner, back from Washington: Elon is “truly an N of one” — 20-hour days, team sleeping across from the White House — and a systems thinker whose first question was “who sends out the wires? Can I just get a list of all the wires scheduled to go out?” The Leviathan “convulses” because nobody looks at the wires. The goal is simply returning to the 2019 baseline grown at 2.5% — which balances the budget within this term by cutting ~$1T of COVID-era excess: “Silicon Valley’s gotten fit… government hasn’t gotten fit at all.”
• The mental model: track one is normal reconciliation — filibuster-proof omnibus, Hassett and the House budget committee, something signable in April or May. Track two is cuts by executive authority — the part causing fury. Exhibit A: USAID, $50B/yr — Elon’s Sunday-night verdict: “there’s no apple to be saved, it’s a total ball of worms… the whole thing needs to be shut down,” with Rubio set to become acting director and possible consolidation into State.
• The legal chess: Gerstner expects a Schumer-led suit as early as this week under Article 1, Section 99, Clause 7 (power of the purse) — but thinks it’s “on pretty weak footing” against the doctrine he called “empowerment”: the president faithfully executing every law, just for less money, with an overriding duty to protect Americans from bankruptcy. Two outcomes: political pressure forces bigger cuts into track one — because “you do not want to be defending every single line item to the American people,” which DOGE’s radical transparency guarantees — or SCOTUS recognizes expanded executive power under that doctrine. Either way, expect “Elon causing a constitutional crisis” headlines.
• The market payoff if it works, per Gerstner’s Sunday tweet: a believable $1T-cut, balanced-budget plan and “interest rates are going to come down” — the bond vigilantes shorted on the assumption of stimulus with no cost discipline, and “people are wrong.”

10. The Golden Age of uncertainty — and a sovereign wealth fund coda

• The synthesis: three stacked uncertainties — technological ($150B companies challenged “by a Chinese startup on a shoestring”), economic (free trade, a career-long constant, now in question), political (a looming branch-vs-branch Supreme Court fight) — force the discount rate up and multiples down. Gerstner tells his team to downsize risk (“you don’t do the tenth best idea”), expects six months of heavy volatility, and notes VC marks lag but follow: “the multiples they pay, that rolls downhill.” Yet fundamentally “the super cycle is on fire” — DeepSeek unleashes inference demand, deep research unleashes more. “It may in fact be the Golden Age and off to the races — but at the moment it’s the Golden Age of uncertainty.”
• Gurley’s cautionary loop-closer: Nvidia bulls who countered the (misread) $6M number by insisting DeepSeek had far more GPUs “may cause some in Washington to say we should have had higher restrictions… you may end up with sanctions that end up hurting Nvidia’s revenues. It’s a dangerous place to play.”
• On the sovereign wealth fund, Gerstner takes both sides: monetizing spectrum and drilling rights for citizens is attractive, but with $40T of debt plus ~$50T unfunded liabilities at 5%, the fund’s hurdle rate is 5% — “we’re levering up the balance sheet of the United States to earn the spread” — plus crony-capitalism risk. “A close call… but I think it’s going to happen either way.” Gurley is “way more skeptical” — the fund could be vulnerable from administration to administration, and “the most successful sovereign wealth funds in the world are all in autocracies.” Gerstner disagrees (Norway, Korea, Canada) but concedes those share independence from unilateral executive control — and recommends Billion Dollar Whale on Malaysia’s fund for how it goes really wrong.