Frontier Labs Threatened by Kimi? Should the US Ban Chinese Open-Source Models & Stripe Buys PayPal

Summary

• China shipped two near-frontier open-weight models in a week — Kimi and Qwen — and the panel’s verdict is acceleration, not surprise: half of OpenRouter traffic already runs through China-created models, and “in a year it could be everybody.” Jason’s caution — “an eval is just an eval” — but the demand signal is real: Kimi is blocked to new consumer signups, and Kimi K3 at ~2.8 trillion parameters is frontier-scale compute, not a laptop toy.
• On banning Chinese models: a blanket ban would be “massive overkill” (Rory), and the anti-ban lineup of Bill Gurley, likely David Sacks and Emil Michael is the tell — but Jason draws a hard line on data-export risk (“you’re not going to convince me… these are companies arguably run by the PLA”) and expects tighter constriction than laissez-faire. The Chinese administration, meanwhile, is talking about banning the sale of those models to the US — “kind of totally zany.”
• The billion-dollar question: does open-weight competition push OpenAI/Anthropic growth sub-100% within one to two years? Jason: “I genuinely don’t know. And if I did, I’d be trading that stock… you know the answer to the entire direction of the US stock market for the next two years” — all hyperscaler RPO hangs on foundation-model commitments. Harry: “if they price Fable at Sonnet rates, I think they’ll own the market” — but the training-cost nut, subsidized by venture capital, means “they can afford to compete 50%.”
• Nobody could explain why America has no open-weight champion while Chinese open-weight players fetch $50–70B valuations. The candidate answer: “the dirty little secret is a lot of their advantage is distillation, which you can’t legally do if you’re US-based” — leaving an ~80%-cheaper intelligence market unserved. Sign of the froth: Harry was offered a Kimi SPV at $20B that morning.
• OpenRouter should sell: “the perfect outcome is to sell the moment it becomes commoditized but before everyone fully realizes it… my gut is it might be now” (Jason), with routing being embedded everywhere (Ramp, Databricks) and hyperscalers natural buyers. His founder math: turn down a $6B offer only for 10x conviction, never 3x — “this isn’t clearing the prep stack. This is clearing my life stack.”
• Infra still beats apps: Fireworks passed $1B ARR in three and a half years (40 trillion tokens/day, mid-30s margins expanding, $1.5B round), while the app layer is “almost a rounding error” — all apps ex-Cursor likely sum to less than what OpenAI and Anthropic spend on training data (Mercor ~$2B, Surge $3B ARR). Jason’s labeling experiment: a tuned model beats the “sea of mediocrity” generic LLMs by an order of magnitude, so generic-only app usage “declines to prototypes.”
• The billion-dollar question: will open weights dent the two leaders within one to two years? Jason: “I genuinely don’t know. And if I did, I’d be trading that stock… you know the answer to the entire direction of the US stock market for the next two years” — all hyperscaler RPO hangs on foundation-model commitments. Harry: “if they price Fable at Sonnet rates, I think they’ll own the market” — but the training-cost nut, subsidized by venture capital, means “they can afford to compete 50%.”
• Stripe–PayPal is “already pre-scripted”: the board rejection is the dance, a 28% premium lands at the mid-30s take-private average, and Advent parks it through antitrust. Strategically ballsy — doubling footprint by buying a 7% grower at sub-10x profits — but blended growth falls toward ~18%, “below the Mendoza line,” and the PayPal mafia’s early Stripe money buys its old company back.
• Venture pricing inversion: hot Series As at $300–500M on $2–5M revenue versus Series Bs at $1–1.5B on $100M+ — “a 3x price increase for a 50x revenue increase” — has made growth the risk-adjusted seat, but Sequoia-style tranche rounds (“nature abhorring a dollar left on the table”) are how these trends end. Adjacent macro call: DRAM makers saying they’ll make 80% operating margins in a commodity — “when that pricing breaks, it’ll be brutal to the downside.”

Deep dive

1. Kimi and Qwen change nothing — they just accelerate everything

• Jason’s opening caution on the week’s two near-frontier Chinese releases: “an eval is just an eval” — don’t take engineers-tweeting-benchmarks as proof; “let’s prove it in the field.” But the demand signal is undeniable: Kimi is blocked to new consumer signups because demand is “I don’t know whether it’s geometric or exponential.” And the deeper point — it’s not new: half of OpenRouter traffic already runs through China-created models, often US-hosted and for all intents open source; “in a year it could be everybody… that is material.”
• David’s read: exactly the trend line — five main Chinese LLM companies, aggressively funded with smart engineers, “6, 9 months behind depending on how you measure it.” His nuance: don’t lump the models together — some DeepSeek models run on a laptop, while Kimi K3 is a ~2.8-trillion-parameter “huge honking thing” needing “myriads of GPUs,” comparable in compute to US frontier models. “After 3 years of competent execution along a pretty defined trend, we now have 3 years and 3 months.”
• The demand backdrop via Jesse Zhang’s (Decagon) data: one highly regulated customer’s token use is up ~2.5x since January, driven by supervisor models tracking agents run in parallel — “four times the agentic use just to have multiple agents regulating agents.” Jason’s conclusion: “the quest for equivalent models at a cheaper price is just going to keep going up.”

2. The ban debate: strange bedfellows, and Jason’s data-export red line

• The firestorm: Dean Ball (likely) — OpenAI’s new policy hire, two weeks in, ex-Trump administration — hinted at restricting Chinese models, invoking “AI communism.” David’s verdict: hysterical in tone and naive in position — when the maker of the $10–20 closed product suggests banning the $2 alternative, “you’re not talking as a common citizen. You’re talking as the provider of the company who will jack up our rates the minute the stuff gets banned.”
• The anti-ban lineup is the tell: likely David Sacks (“this is rubbish, stop”) and Emil Michael — a man who “knows how to hate,” whose longest-running hate is Bill Gurley — landing on Gurley’s free-market side. “If Bill and Emil are on the same side saying don’t ban these models, you got to know there’s got to be some truth in that.” David’s line-drawing: a blanket ban is “massive overkill,” but the White House will not run on Kimi “even if it’s Kimi hosted in California” — while a Decagon doing boring customer-support inference has “no reason why you should pay marquee prices when something 10x cheaper is available.” Meanwhile the Chinese administration talks of banning sales of those models to the US: “kind of totally zany.”
• Jason’s dissent, kept whole: “I don’t think you’re going to convince me there aren’t some data export risks with China-based models… we cannot understand what these models do. They are connected to the internet.” These are companies “arguably run by the PLA,” and a Fortune 500 CIO with their job on the line won’t accept internet reassurance that on-prem removes the back door. His call: tighter constriction than laissez-faire, even without a ban.
• Rory’s rebuttal — technical, but conceding the politics: the risk is real (Boeing suffers continual state-attributable cyberattacks; Huawei was shut out of Western cellular on “as yet unprovable fear”), but these are open-weight, not open-source — you see weights, not training data — and a trusted US inference host could plausibly block exfiltration. “In logic you could satisfy a technologist that the risk is not there. Whether you can satisfy a politician… is another question.”

3. Nobody can explain why America has no open-weight champion

• Rory’s core question: “is the open-weight low-cost LLM business a good business? And if it is, why can’t some red-blooded American company step up and give OpenAI and Anthropic a run for the money?” Chinese open-weight players fetch $50–70B valuations — “I wouldn’t turn down a $50 billion outcome.” Where’s likely Grok, Gemini, likely Llama, Reflection, Thinking Machines — which just announced Inkling, positioned as something to build on rather than frontier?
• The candidate answer: “maybe the dirty little secret is a lot of their advantage is distillation, which you can’t legally do if you’re US-based.” Dissociate the bundled frontier product into IP plus inference and the open route is roughly 80% cheaper intelligence. Harry: with mass demand for that, when does a US company fill it? “I’ve asked so many people… no one’s actually given me an answer.”
• Harry on the scale of the puzzle: a new category with two premium companies at ~$2T combined market cap on ~$100B+ combined revenue, four or five capable US builders shipping nothing, and five Chinese companies “cranking night and day.” Froth check: Harry was offered a Kimi SPV at $20B that day — “we’re oversubscribed, but we’ll make room for 5 million for Harry.”

4. OpenRouter should sell into the commodification window — but only a founder decides

• Jason on why leaking the sale talks was “very savvy”: OpenRouter was early to a heterogeneous-model bet that “probably seemed too nerdy and too niche” at the start, but routing is becoming an included feature everywhere — Databricks’ gateway, now Ramp’s competitor. “The perfect outcome is to sell the moment it becomes commoditized but before everyone fully realizes it… my gut is it might be now.” Commodifying spaces don’t kill everybody, “but it might maim you” — “probably why Cursor wasn’t dumb to sell at 60 billion.”
• Harry on the buy side: standalone NPV is maybe “a couple of billion, not huge” — but to a hyperscaler with no single in-house model to push (Amazon especially, “maybe Microsoft now that the divorce is coming true from OpenAI”), owning the model-agnostic layer could shift 10% of enterprise share over half a decade. The tell: “I remember thinking, damn, I wish I was in that — which is always how you know what a venture guy really thinks.”
• Jason’s founder math on Harry’s “would you sell at 6 billion?”: if it’s purely financial, sell — but to turn down a nosebleed offer “it has to be 10x to go for it. It’s not worth it for 3x.” Going from $40K in the bank to $400M versus $800M “is irrelevant if there’s risk,” and the next three years are “a handful of sweat and a handful of market change and a handful of people that quit.” “This isn’t clearing the prep stack. This is clearing my life stack.”
• Rory’s board craft: the private-liquidity window “opens so rarely that it’s always a good idea to pay attention” — but the founder decides, since even with board control, 70% ownership and a drag-along, “if the founders who are core to the business don’t want to sell, it’s not going to happen.” His translation of Jason: “if you turn down a big-ass offer, you better be sure it can be way bigger. High certainty and high bigness.”

5. Fireworks and the inference boom: expanding margins now, capex reckoning later

• The round: $1.5B from Index, Gavin Baker, Nvidia, Lightspeed and 20VC (Harry cited a $17.5B valuation), over $1B ARR in three and a half years, and 40 trillion tokens a day, up from 15 — with Lin expecting to double by year-end. Rory: Fireworks, likely Baseten, likely Fal and Together are the one-to-one beneficiaries of the open-weight explosion, because “what you’re not going to do is be using the API back to China, even if they’d let you.”
• The margin flip: on paper a compression story — buying neocloud compute and reselling it with hosted LLMs — but with demand this massive, “whatever compute you own now, you can charge way more.” A lowish-gross-margin business is suddenly growing ~5x to a billion with expanding gross margins; Lin put it at mid-30s, moving up as they eat more of the stack.
• Rory’s caveat on that same plan: eating the stack means vertically integrating into data centers — “a ton more capex” — so the commodification risk is deferred, not gone.

6. Apps are a rounding error; everyone will end up training their own model

• Jason, the infra bull despite himself: “all the good investments sure seem to be in the infrastructure… I’m waiting for the era of the application layer in AI… but I don’t believe it’s here yet.” Vibe coding “didn’t kill software, but where is the software renaissance?” App revenue is “so trivial compared to the infrastructure. It’s almost a rounding error.”
• Rory’s three buckets: infrastructure spend at $800–900B a year; the two foundation-model companies at ~$100B; and every other apps company rounded up — “you struggle to make 40 or 50 bill,” starting with Cursor at ~$4B. The kicker: the sum of all apps ex-Cursor is “probably less than the amount that Anthropic and OpenAI are spending on training data” (Mercor ~$2B ARR, Surge $3B, Handshake $1B). “At some point, the people spending a trillion dollars a year are going to want some apps to pay for all this.”
• Jason’s hands-on epiphany from labeling data for his agentic recruiting app (built on Sonnet and Opus): hand-labeling made it “exponentially better” — “the outputs are literally an order of magnitude better once you do it.” Generic LLMs are “a sea of mediocrity combined into one giant LLM… every mediocre history professor, every mediocre doctor that doesn’t even know what caused your runny nose”; 20–30 questions from a true domain expert is “a step function.” He’s stopped mocking the labeling market: “I ain’t making fun of it anymore.”
• His structural call: at scale every application company wants its own model, and generic-model-only usage “is just going to decline to prototypes” plus state-of-the-art for the expensive parts. Rory’s discipline on what that means for Mercor et al.: bullish only if enterprise custom models are additive to foundation-model growth — if open weights instead dent OpenAI/Anthropic, “worrying about your Mercor valuation will be the least thing people are worried about, because you’ll see an implosion of much bigger market-cap entities.”

7. The billion-dollar question: the OpenAI–Anthropic growth rate IS the market

• Harry forced a yes/no: will open weights dent the two leaders within one to two years? Jason: yes at some magnitude, but the real question is whether growth falls sub-100% — “I genuinely don’t know. And if I did, I’d be trading that stock… if you know the answer to that one question, you know the answer to the entire direction of the US stock market for the next two years,” because all the hyperscaler RPO is a function of foundation-model commitments. If 10x slips to 2–3x, capex gets repurposed toward inference, “but it will be a big-ass dislocation.”
• Harry’s evidence that competition already bites — pricing behavior: Fable “went from you can’t use it, it’s not secure,” to government-approved, to now 50% of your whole monthly usage — why loosen access when capacity is short? “Competition, right?” His call: “if they price Fable at Sonnet rates, I think they’ll own the market” — the constraint being a high cost base “subsidizing it with venture capital.” “They’re pushing it as far as they can… they can afford to compete 50%.”
• Rory on why a price war is harder than software’s bundling wars: Microsoft crushed rivals with zero COGS; models carry real serving costs plus training recovered “pretty damn quick, because it only lasts 12, 24 months before it’s obsolete,” and a 20x-revenue valuation demands Microsoft-like 40% operating margins. The grace: “if you’re growing 10x year on year and you have any kind of positive and improving gross margin, it just covers all the nut” — but if price cuts become the only way to hold growth, margins deteriorate and “that in itself would be a different ballgame.”
• The Nvidia coda: Harry can’t understand why Kimi doesn’t lift his position (“how long are you going to stay flat for?”). Jason’s three scenarios from the $180–210 range: capex elevated-but-not-doubling → the stock grows into the valuation; another step-up like “the Claude lift” early this year → the next leg; any slowdown → “even this valuation will look crazy.” Gavin Baker’s cross-sectional rule: whatever assumptions value Nvidia must, to a rounding error, value DRAM and every other bottleneck. “The only thing that matters is the OpenAI and Anthropic growth rate in ‘26.”

8. Stripe–PayPal is already pre-scripted: 28 goes to 35

• The shape: both process ~$1.8–1.9 trillion a year, yet Stripe is valued ~$150B while the joint Stripe–Advent offer for PayPal is ~$50-something billion — PayPal trading at sub-10x profits, ~1.7x revenues. Adjusting Stripe’s net-reported ~$6B against PayPal’s ~$30B gross, PayPal is still ~1.5x Stripe’s size like-for-like. Rory: “a big ballsy play to double your market cap” — Dell-like, but Stripe must do “a lot of hard-nosed stuff” to a company that’s been “a revolving door of executives” since the PayPal mafia walked out.
• Jason’s stressor: blending a 7% grower into a 20–30% grower drags you toward ~18% — “below the Mendoza line of 20% growth at scale” (Rory’s pushback: “there’s no such thing as a Mendoza line for growth at 5 billion and above”). On integration, Jason’s learned comfort: you never actually rationalize the two spaghetti codebases — “you fix it over five years, or you have an LLM lift.”
• Both call it done. Jason: the board rejecting it “means to me that they’re going to accept it” — no bank lets you lead with your best offer, and a 28% premium against a mid-30s take-private average is “the perfect amount of back and forth. 28 to 35. It’s already pre-scripted” — then parked with Advent while antitrust and capital get sorted. Rory’s Delaware walkthrough: the board can only refuse under business judgment if standalone credibly beats the bid — unless PayPal’s five key internal metrics are already turning, an independent director on $300K of RSUs isn’t playing hero.
• The poetry: early PayPal mafia money — Peter Thiel most notably, “along with Sam Altman’s 2%” — sat in Stripe’s earliest rounds and now buys PayPal back 15 years later. “The likely Padawan finally becomes the Jedi… they’re getting the old gang back together.”

9. Growth has been the cheap seat — and tranche rounds signal the late innings

• Harry’s market read: Emergent (AI coding, $120M ARR) raised a $130M Series C at $1.5B post in July, while hot Series As price at $300–500M on $2–5M revenue — “a 3x price increase for a 50x revenue increase.” Rory’s risk-adjusted comparison is to pay ~4x more “for incredible PMF and 70 to 100 times revenue scaling” (the Factory deal at $1.5B versus its earlier $300M round) — or Fireworks at $17.5B against a claimed $2B year-end run rate, under 10x.
• Jason’s structural point: below growth stage “you better be a damn good picker” — Series B now demands seed-investor skills where it used to be math and team assessment. “You don’t really want to be a picker. You want to be a pricer.”
• Rory agrees, with a cycle warning: since 2022 growth has been carried by momentum — “at least 40% of the unicorns minted in Q1 or Q2 of 2025 will have had a subsequent markup” by end of Q2 ‘26, “good things get more good things” — and Sequoia-style tranche rounds are “nature abhorring a dollar left on the table,” pricing the excess return “away from Harry and back to us.” “This is how these trends end.” Long-run CAPM says early must out-return late — “and by long term I mean longer than you’ve been alive, Harry.”
• The tranche ick, preserved as mild disagreement: Jason wouldn’t charge one investor $1B and another $5B in the same week — “80 to 90% of a good deal to me always seemed to de-stress my life” — though at massive size “I have to say yes to that as a founder.” Rory: icky but survivable; real structure or heavy debt are worse cap-table sins, and it’s fine “provided you don’t give a damn that they’re second-class citizens.” Harry’s counter stands: a $5B Sequoia-led headline “will create fear among other VCs to fund competitors.”

10. Trusted supply chains vs the DRAM knife fight — and a weird public/private world

• Rory’s contrast of two adjacent chains riding the same AI wave: Nvidia–TSMC (“famously they don’t even have a written contract” across 30 years) and TSMC–ASML raise prices gently and plan in decades, while DRAM’s three players — the two Koreans and Micron — play “screw you, we’re raising prices 40% this quarter… we’re going to make 80% operating margins in what Harry would call a commodity, because we know that two years from now you’re going to screw us.” The tradeable aha: “when that pricing breaks, it’ll be brutal to the downside — but maybe that’s a year, two years from now.”
• CoreWeave “depressed for a long-ass time,” per the panel: memory prices have doubled the cost of building the product, the big OpenAI commitment concentrates the story, and public-market gravity sets in — though Rory still calls it “attractively valued on a sales multiple basis.”
• The backwards world: three next-gen nuclear companies SPAC’d public and swing 50% a day (while Valar Atomics — likely — raises privately at a 3x step-up in four or five months), yet cash-flowing Stripe and Databricks ($3B Series M at $188B) stay private — “the SPACs are taking stuff public that should probably be venture-backed, and the very best venture assets are staying private long after they’re kicking off cash.” Footnote from the likely C-Square data-center IPO — $1B run rate, 16% growth, $3B cap, no AI multiple: “you got to deliver… the market may be exuberant… but it’s not stupid.”