China Open-Source, Compute Arms Race, Reordering Global Trade | BG2 w/ Bill Gurley and Brad Gerstner

Summary

• Groq COO Sunny Madra’s core call: China’s seven-plus open-source labs are compounding on each other via distillation — K2 is “sort of a well-known remix of what DeepSeek had done” — and a Qwen 30B-parameter release now performs “as good as GPT-4o.” The Chinese leaders offer “90% of the quality in terms of intelligence, but at a 90% price discount,” and enterprises worldwide are taking that trade over American-values alignment.
• Yet Sunny predicts the pendulum swings back: by Q4 this year or Q1 next year a top-three worldwide model will be US-based open source, driven by OpenAI’s imminent open-source release (enterprise demand is “like a Tesla Roadster… that’s the model that everybody wants to use”) plus Meta’s efforts. Enterprises want brand and accountability — the Red Hat-on-Linux pattern — so a US-domiciled equal on price/intelligence would “rise to the top in a few days” on OpenRouter.
• Compute demand shows no evidence of a glut. Google went from 5 trillion tokens/month to a quadrillion — “that’s a thousand trillion… 200x” in a year and a few months; Groq racks are “fully consumed within a few hours” of firing up; Elon is targeting 50M H100-equivalents (~11GW) and the Abilene deal is 4.5GW, “well above” the $500B pledge. Brad cites CNBC’s report that Iconic is expected to lead a $5B Anthropic round at $170B; xAI rumored at $150–200B — OpenAI + Anthropic + xAI would be roughly $1T combined.
• Bill Gurley’s caveat on the boom: “no one here is raising price… everyone’s pricing to share. That means they’re pricing under cost” — negative gross margins are rumored at top AI brands, and Anthropic throttles usage rather than raise prices. When the Oracle/Sun → Linux/MySQL cost-optimization phase arrives, “that will create a bump in the supply-demand curve.”
• The model layer is commoditizing; value accrues at the application/consumer layer. Brad: ChatGPT may cross 1B weeklies this year and owns the first consumer threat to Google in 20 years; Bill argues OpenAI’s durable moat is “switching cost and lock-in more than… staying on the edge of the model race.” Coding agents are already distilling Apache-licensed Chinese models to “eliminate the entire cost of Sonnet… for 70% of use cases.”
• On tariffs, Brad declares the Bessent ~$300B consensus won over “nuclear Navarro”: EU at 15%/0% plus ~$750B energy purchases, Japan at $550B directed investment, Brad recalled Bessent saying June had the first monthly Treasury surplus since 2015, import prices rising slower than domestic goods, Atlanta Fed GDPNow back at 3%. “All we’ve seen so far is deals deals deals deals… he’s in line for a bonus” — hedged with “yet” on inflation.
• Brad goes “out on a limb” against consensus: a very big China deal gets done before year-end — the president is “a deal junkie,” not a China hawk; Beijing postponed retaliatory tariffs and invited a September–November visit; the deal could span rare earths, chips, “maybe even military cooperation,” with Brad expecting China to pay at least 15%. On positioning: back in since May 2, up 30% off the NASDAQ bottom, but “I think we’ve captured a lot of the return for the year” — still bullish on AI opportunities, including Groq’s (accidentally disclosed) new round.

Deep dive

1. China’s open-source flywheel: distill, remix, compound

• Sunny’s mechanism for why China is accelerating: open weights let seven-plus deep-pocketed labs “compound on each other” instead of building giant training clusters in silos — “almost consider it like a remix of someone’s model. K2 sort of a well-known remix of what DeepSeek had done.” Two dimensions moving at once: frontier models improving faster, and rapid distillation into small “turbo” models — a Qwen 30B-parameter model released that day performing “as good as GPT-4o,” which “was world class not that long ago.”
• The IP angle Sunny flags from the AI summit: the president addressed copyright with the reading-a-book analogy (“if you read a book and you use it, you’re not violating the copyright”) — while China simply “worked around that because of their position on IP.” Bill adds the 20-year backstory: when “most of the world accuses you of IP theft,” embracing open source is natural — though he admits he doesn’t know “was the government promoting open source… or did it just develop through competitive forces?”
• Bill’s farming-community analogy, kept as told: two communities of 10–20 farms each; one just competes at the weekly farmers market, the other is forced to also share all best practices. Run it two years and the sharing community has higher global output — “a higher fitness level for a community” — though “a lot less chance of a breakout monopolist like we’ve had in many of the sectors in American technology.”
• The pace is humbling even insiders: Bill “almost feels like an idiot” — Zhipu (likely) releases a model, he looks it up on PitchBook, and they’ve already raised $1.4B. “They’re coming out of everywhere,” suddenly atop the OpenRouter leaderboards. Meanwhile reasoning models changed the game itself: post-o1 (Noam Brown’s program), models don’t compress the whole internet — they use tools, and “none of the benchmarks really allow you to use a tool,” so tool-using capability is not captured by current benchmarks.

2. 90% of the intelligence at 10% of the cost — and demand follows

• Sunny’s chart: Chinese open-source leaders cluster top-right on intelligence-vs-price — “90% of the quality… but at a 90% price discount. Whenever you offer that to anybody, you’re going to see people want to use that.” Groq is seeing exactly that demand, from Saudi Arabia to individual developers.
• Brad’s blunt read on the values debate: “if you deliver something really powerful and really cheap, that’s more important to these players than American-values-aligned.” But give the market something equally powerful, equally cheap and Western-aligned, and Sunny says it wins — “on a place like OpenRouter… we would see it rise to the top in a few days.”
• The Red Hat logic for why US brands still matter: enterprises “want some accountability… someone that they can go and point to if they needed something.” Legal sign-off, liability, risk — “if that shows up, it will win,” triggering “a huge shift back towards those models versus the Chinese ones.”

3. The American open-source counterattack

• Sunny’s dated prediction: “Q4 this year or Q1 next year… a top-three worldwide model will be a US-based open-source model” — combining OpenAI’s open model (Sam has flagged a late-summer release) with Meta’s efforts. Enterprise anticipation for OpenAI’s release is unlike anything he sees: “it’s almost like the demand for a Tesla Roadster… everybody asked for it.”
• Bill’s Alibaba insight generalizes to the US: asked why Alibaba funds competing model startups while having Qwen, his answer — “if you’re not confident you’re going to win on offense, you want to play defense”; commoditizing a potential threat (no ByteDance-equivalent in AI) is rational. Corollary: laggards like Microsoft, Amazon, Apple “should be funding an open-source competitor” rather than, e.g., Amazon backing Anthropic.
• Bill predicts new US entrants that co-evolve with Chinese models — “is Linux American? Is Linux Chinese? No one thinks of it as having a domicile” — offering sanctioned, “cleaned” Red Hat-style versions cheaply, followed by “a big fight around regulatory capture.” And the kicker: “if I’m the Mistral team, if you’re not distilling on these Chinese models, I don’t know what you’re doing right now.”
• On Meta’s rumored open-source retreat, Brad’s hunch: “that’s a misread” — they stay committed but complement with a proprietary model, partly because openness is the recruiting pitch (“everyone they’re pulling over are coming from closed places”). On xAI: Elon open-sources one generation back, but Sunny guesses Grok 2 or 3 is now “so far behind that… what’s the purpose?” Bill notes some researchers hold open source as “a religious belief… oddly higher on their Maslow hierarchy than the money.”

4. Compute: a quadrillion tokens and no evidence of a glut

• Sunny’s single best data point against the overbuild thesis: Google went from 5 trillion tokens/month to a quadrillion — “that’s a thousand trillion… 200x” — in a year and a few months. Brad adds: “Every single search query on the planet today is now an inference transaction,” and at Groq, “anywhere we lay infrastructure, we fire up a rack, it becomes fully consumed within a few hours.” Asked directly for any evidence of supply outstripping demand: “No.”
• The scale of stated buildouts dwarfs last year’s debate: Elon’s xAI target of 50M H100-equivalents (Clark Tang’s breakdown: ~4M physical GPUs, ~11GW), the 4.5GW Abilene deal coming in “well above the $500B estimate” promised to the government — all reflecting Jensen’s inference-going-1-billion-x call from the pod last year.
• Anthropic’s throttling announcement is the tell: too many people “using way too much of our services” — Sunny floated “Jevons’ paradox, or whatever.” Bill on his own behavior confirms it: reasoning models silently burn “10 to 100x more” tokens than a first-generation AI search.

5. Sport of kings: pricing for share, not for cost

• Bill has “never seen anything like it” — Uber/Lyft was the precursor, when private companies first proved “they can be more risk-seeking with capital than the public companies are allowed to be.” OpenAI will lose $7B this year competing with a Google that “would never allow themselves to do that.” He dubs OpenAI, Meta and xAI the “cost is no object” group — “the CNO group” — versus Microsoft trimming capex and Amazon not keeping up Nvidia purchases relative to AWS share, with Anthropic ambiguously in the middle. “It’s a sport of kings” — and the picks-and-shovels winners are Nvidia, Dell, “maybe SK Hynix too.”
• His load-bearing caveat: “the thing that throttles demand is price, but no one here is raising price… everyone’s pricing to share. That means they’re pricing under cost.” Even the power-shortage argument dissolves — “you wouldn’t run out of power if you just took the price up.” Rumors of negative gross margin at some of the best-known AI brands; when the market moves from win-at-all-costs to the Oracle/Sun→Linux/MySQL optimization phase, “that will create a bump in the supply-demand curve.”
• The optimization phase is already starting where inference intensity is highest: coding-agent companies are building their own models off open source — distilling Apache-licensed Chinese models to “eliminate the entire cost of Sonnet that sits under it for 70% of use cases,” reserving frontier models as the gold tier above silver/bronze built at “one-tenth the price.”

6. Where the value lands: consumer lock-in, not the model layer

• Bill’s book anecdote as the moat argument: OpenAI “knows a tremendous amount about my book right now” without re-prompting — so OpenAI’s best shot at long-term success “comes from switching cost and lock-in more than it will come from staying on the edge of the model race.” Brad adds the strategy: use the high-gross-margin consumer business to subsidize share in coding/enterprise — “it reminds me a little bit of Amazon… the monopoly retail business subsidizing AWS for a decade and then begin to take price.”
• Sunny’s TPU question — does Google’s vertical integration confer strategic advantage as apparently the largest token processor? Bill’s test: how many non-Google applications run on TPUs? Perception is most TPU volume is Google’s own apps; the advantage only matters if there’s a Google Cloud crossover moment. Brad: too early to know, though OpenAI using TPUs for some inference is a real signal.
• Brad’s synthesis: “the battle for the consumer is ultimately where the value occurs… not who runs what hardware.” Google owned the verb for 20 years; ChatGPT — likely crossing 1B weeklies this year — is the first real threat. If seven Chinese labs can distill each other, drive up intelligence and drive down cost, “the model layer is being increasingly commoditized.” Sunny adds that applications and coding agents will be a heterogeneous, lower-margin world with lots of players. Still, nobody predicted OpenAI + Anthropic at over half a trillion combined, “and you throw xAI in there, it’d be a trillion dollars.”

7. Tariffs: the Bessent consensus won — “deals deals deals deals”

• Brad’s reconstruction of the year: a team-of-rivals White House offered door one (Bessent/Lutnick, 10–20% across the board, ~$300B total vs $75B in 2024) or door two (“nuclear Navarro” — replace the IRS with $2T of tariffs). “Until the president tells us whether it’s door one or door two, we’re out… the market shot first and asked questions later” — NASDAQ down 21% at trough, then a 30% move in 60–70 days.
• The heterodox theory held — so far: Bessent/Hassett argued exporters must eat a 15% tariff or lay off millions, against 90% of economists predicting US consumer inflation. The NEC paper deconstructing core PCE shows import prices rising slower than domestically produced goods — “the exact opposite of what you would have expected.” Brad’s explicit hedge: “I will caveat this by saying yet,” core PCE bottomed and ticked up, and “it’s almost impossible… zero mistakes” in reordering all of global trade on the first pass.
• The scoreboard: EU at 15% in / 0% out plus ~$750B energy purchases; Japan similar plus $550B invested “in a way the president gets to direct”; a $300–350B recurring tariff revenue stream; Brad recalled Bessent saying June saw the first monthly surplus since 2015; Atlanta Fed GDPNow back at 3%. The CEO analogy: if a company CEO had pitched this high-risk plan in January, “measuring them halfway through the year… he’s in line for a bonus” — versus Larry Summers calling it “the biggest economic disaster of his career” months ago.
• Reshoring is now investable: Altimeter just led (announced on the pod) a Series A in an all-American rare-earth magnet producer — “viable investments because of the tariffs.” Sunny’s operator view: Groq’s tariff schedule moved “six to eight times” and Groq is negotiating with component suppliers; planning has settled from day-to-day to monthly — but “we still have a big thing looming with a China discussion.”

8. The China big-deal call — and Brad’s market temperature

• Brad goes “out on a limb” against consensus (which expects the China piece to be a problem or small): “this president wants to do the biggest deal ever done with China… he’s not dogmatic at all… ‘I’m a deal junkie.’” China postponed retaliatory tariffs and invited a September–November visit; Brad expects a deal before year-end that could include rare earths, chips, trade rebalancing, “maybe even military cooperation” — anchored by the president’s line that with a magic wand he’d “cut the defense budget in half for the United States, for China, and for Russia,” which Brad calls “an extraordinarily flexible mindset.” Brad expects China to continue paying at least the pre-Trump 15%.
• On positioning, pressed by Bill (buy low, sell high, at all-time highs?): out in March, back in on May 2 when “the Bessent consensus has won,” up 30% off the bottom, ~10% YTD — and “I think we’ve captured a lot of the return for the year.” But bottoms-up, “we see tons of opportunities” in AI — including, accidentally disclosed on air, Groq’s huge new round (“his face is turning all red”), atop a rumored $600M raise at $6B on a few hundred million of revenue, maybe doubling year-over-year.