Enterprises Fear Frontier Models | Sam Altman Offers Trump 5% of OpenAI | DeepSeek Builds Own Chips

Summary

• The title debate — OpenAI floating a 5% stake for the US government. Rory’s verdict: a catastrophic unforced error — “it’s like rewriting Atlas Shrugged where John G goes to Washington and says why don’t you regulate me more?” Ownership doesn’t buy political alignment (Microsoft owns 30% of OpenAI and they’re “in a stale marriage looking for a divorce”), and once you’ve claimed you’re destroying a $30T labor market, “it’s about an hour before Bernie Sanders says… maybe we should go for 50.”
• Jason’s counter after initially agreeing: Sam Altman is one of the great investors of all time running the Klaviyo-gives-Shopify-5% playbook — small stakes to giant partners create “an unexpectedly large amount of alignment” — and he’s deliberately anchoring at five rather than fifty because the government taking a stake may already be effectively decided (see Intel).
• Frontier AI has entered the permission economy: Washington’s 19-day Fable 5 ban ended with discussion of a structured pre-approval process that hasn’t been finalized. “Six months ago you could ship software like a free man.” It’s the exact inversion of the 1990s internet gifts (telecom dereg, Section 230, no sales tax) that bought Silicon Valley a 20-year leave-us-alone run — “we sneered at GDPR and here we are.”
• The compute money machine is entering its late-cycle phase: Meta jumped 10% for selling compute it failed to use, and Nvidia’s “compute now, pay later” recognizes hardware revenue upfront with put-back rights — “as legal as church on Sunday” but a derivative bet on demand never slowing. The warning: “the time to manage for the downside is when no one is managing for the downside” — and per Jason, nobody is. Meanwhile Anthropic (with Samsung) and DeepSeek are building their own chips — not for customization, which one speaker calls “soft language,” but to recapture Nvidia’s margin.
• Enterprise adoption, not model quality, is the bottleneck: Karp’s CNBC claims were “spot on” (ROI doubt plus data-trust fear; Palantir rose 9% that day), and Microsoft’s $2.5B, 6,000-person forward-deployed-engineer push is the response. Harry: “it’s going to fail” — the talent depth doesn’t exist. Rory: the real variable is diffusion rate — OpenAI and Anthropic hit $12B and $4B revenue faster than any technology ever, and “if the next 10x takes three times longer because corporate America can’t adopt, that’s going to have consequences.”
• China is running the counter-strike: Kling raised $2.8B at $18B on ~$500M Q1 ARR — “the most commercially successful AI video product on earth is Chinese” — and the top six models on OpenRouter are Chinese open source. Jason, back from two weeks behind the firewall where even Hong Kong can’t access ChatGPT or Claude: “what do you expect when you can’t access the leaders? They’re going to build something as good or better.”
• Tokens versus dollars: frontier models keep the money even as usage goes open source — “all the tokens can be in one place, but all the dollars can be in the other place.” Jason’s proof: 10 hours and ~$500 stuck on a Sonnet-plus-open-source mix in Replit, solved in 20 minutes with Fable and Opus — cheaper, not just faster.
• Venture’s new physics: founders no longer fear dilution (Ramp at ~24 rounds, Dario at 1-point-something percent of Anthropic) or high-priced last rounds, so Rory now doubles his mental entry price twice — “a seed deal at 60 is really at 240.” And employees are “single-shot VCs”: tender offers are the new IPO proxy, so the trick is joining the company that starts tendering right after you vest.

Deep dive

1. The permission economy arrives — “ship software like a free man” no more

• Rory on the lifted 19-day Fable 5 ban: relief wrapped in a quagmire, because the industry is now “entrapped in some kind of pre-approval process” that hasn’t been finalized. The zoom-out: “six months ago you could ship software like a free man and now you have to get permission from Washington.” Part of why the US economy is so dynamic is that it lacked exactly this — “we sneered at GDPR and here we are.”
• Jason’s deflation: whether more regulation is better is “beyond my scope,” but this oversight is simply “the grown-up state of LLM and AI” — and this particular ban is minor because Fable moves to variable per-token pricing in a week or two: “most of us aren’t even going to use it because it’s too expensive… a niche model until it percolates into the standard Opus and Sonnet.”
• Rory’s history lesson on the inversion: three 1990s gifts — telecom deregulation, Section 230, no sales tax — bought the internet a 20-year “leave us alone” run. Now the posture is “don’t miss us, regulate us… pick us.” “Oil and gas must be looking at this going wow, these people are crazy — no one down at Exxon is saying why don’t we give Washington 5% and check in advance before we do drilling.”

2. Sam’s 5%: fool or genius

• Rory’s case against, in full: the offer solves none of the vaguely credible security issues — it flows from OpenAI’s nine-point plan to restructure US taxation (more on capital, less on labor) because AI will destroy jobs. But 5% of Anthropic is $50B against a Congress raising ~$5T a year — “your whole donation gets rid of 1% of a raise for one year.” Meanwhile OpenAI has “been lapped by your direct competitor” and is lobbying Congress instead.
• The escalation logic: if you’ve told Washington you’re destroying labor in a $30T economy, “it’s about an hour before Bernie Sanders says… maybe we should go for 50” — $140 a head won’t “keep the wolf from the door.” Nor does ownership buy alignment: Microsoft owns 30% of OpenAI and they’re “in a stale marriage looking for a divorce,” and TARP showed the government takes a stake and then “tells you who you can pay.”
• Jason’s counter, after starting where Rory is: Sam is “one of the most successful investors… of all time” running OpenAI like a super-startup, and portfolio experience says the Klaviyo-gives-Shopify-5% move works — “it creates an unexpectedly large amount of alignment… I’m constantly shocked how much that brings you into the boardroom.” As an investor, “I take the dilution.”
• A sharper read: Sam is anchoring at five rather than fifty — “maybe it doesn’t matter what we think because it’s already happened,” with the government already taking stakes (Intel) and Sam telegraphing endpoints early. Rory’s why-they-do-it: the catastrophe narrative is what raised the billions — “you needed to tell a story like every great CEO” — and once you believe it, this becomes “next-level logical.” Note the proposal said companies should give 5%: “he’s volunteering other people’s capital,” Anthropic included.

3. The dilution-insensitive age

• Jason’s observation: massive dilution “has been institutionalized” — hot startups run 16-20 rounds (“Ramp’s done 12 announced rounds according to Claude… I’m going to guess it’s more like 24 with stub rounds”), and at Anthropic, “the most successful startup of our lifetimes,” Dario holds 1-point-something percent and Sam nominally zero — which “takes the edge off the dilution conversation because it’s someone else’s money.”
• Rory’s confession on his own math: dilution has doubled his mental entry price twice — “a seed deal at 60 is really doing it at 240. That’s the honest math today.” Spark will deservedly own ~1% of a huge winner; SpaceX’s ballsiest early check sits at 3-4%. The Carta data softens it: dilution per round is falling, so more rounds can net out to the same dilution.
• The other dead fear: “no one’s worried about making their last round high-priced investors money anymore. Literally no one is” — investors have learned to take 1x “without drama, without blocking, without threats.” Jason: “I was terrified as a founder… I would get blocked by the douchebags. I just don’t see any of that fear exist in founders anymore” — and it adds velocity.
• The Linear counterpoint: Karri has raised twice and refuses all VC intros. Harry loves it (“it will return my fund one multiple times over”), but Jason hears “a little bit of Brian Armstrong in that” — was capital efficiency the right choice in 2026 if the prize is $100B? Rory’s frame: optionality versus upside — if skimping cuts a trillion-dollar shot by even 10% it’s a huge mistake; if the prize is $1-5B, over-raising kills the life-changing exit (“20% of a billion… especially with QSBS”).

4. Karp was right — ROI doubt and the data-slurp fear

• Jason watched the supposedly deranged CNBC appearance and found Karp “more stable than he normally does”: strip the personal baggage (calling Dario a “world historical figure” — a Hegel concept from a doctor of German philosophy) and both claims land. Corporate America is asking “am I getting anything?” — and “am I giving them all this information… are they going to be selling my business to everyone else?” Self-serving, since Palantir sells the cure, but the stock rose 9% that day.
• Jason’s tie-in: the training fear looks exaggerated on today’s terms of use, but vendors will push limits — “OpenAI, Anthropic kind of lied about the books and definitely lied about training on YouTube” — and HubSpot just walked back pooling customers’ prospecting data within a week of announcing it. “Every vendor seeing massive competition or slowing growth is going to be tempted to cut corners on training privacy.”
• The positioning point both liked: Anthropic had “opinions” about how the DoD should use its AI, and Karp pounced — “when people are giving you millions of dollars, they don’t want your freaking opinions, they want your technology.”

5. Meta’s neocloud pivot — plan B gets a 10% pop

• Meta will sell excess AI compute, hosted or raw GPU-by-the-hour like CoreWeave or Nebius; the stock jumped 10%, its biggest single-day gain in five months. Jason: “why not earlier? Didn’t bother Amazon opening up AWS back in the day.” Rory’s stranger reading: two companies (SpaceX the other) bought compute for proprietary assets, failed to build them, pivoted to selling it — and both were rewarded. “It is odd to be able to have a plan A, reverse it, go with plan B, and get a 10% lift.”
• Rory’s scenario tree: either the market believes Meta has a wonderful long-term AI use it can’t yet articulate, or it’s saying “both of you have failed at your long-term goal, but being a cloud provider is a great business — go team.” Two more entrants into a crowded market made the 10-15% decline in Nebius and CoreWeave exactly right. The risk: if more companies discover they overbought, you’re left with few buyers and many sellers — “maybe it won’t be such a good business two years from now.”
• Jason on execution: asking whether Zuck can pull it off is the wrong question — “if a customer wants to buy something, it doesn’t take a freaking genius to sell it if you have it.” The only variable is whether another five gigs of demand exists. Meanwhile Zuck paid ~$900M (Jason thinks well over $1B) via the CRED investment to make Kunal Shah head of WhatsApp — “treading water while he figures it out,” which Jason endorses: “when the core is so successful, stay in the game.”
• The board-member test both land on: AI-improved ad targeting is real but doesn’t justify ~$70B of spend — yet with a $100B cash-flow engine there’s no fatal-error risk: “you’ve earned the right to play. Worst case we spend 70 billion and we’re wrong, just like VR.” Rory’s big aha: capex won’t stop from the supply side — “the demand side is going to be what shuts off the spigot,” and with OpenAI/Anthropic revenue still 2x-3x-ing, it’s wide open.

6. Nvidia finances its own demand — and its best customers tape out

• “Compute now, pay later”: Nvidia sells chips to next-generation neoclouds, recognizes the hardware revenue upfront, and grants put-back rights if the buyer can’t use the compute. Rory: “as legal as church on Sunday, it’s ASC 606… but it is pretty aggressive” — real contingent liability, and if demand slows Nvidia isn’t just growth-challenged, it’s “debooking prior revenue.” Jason’s snark: “now that roundtrip revenue is totally cool and not something you go to jail for — let’s do it every single place we can find it.”
• The cycle tell: the pool of customers who can pay cash “is tapping out,” so growth requires subsidizing new ones. Rory, who publicly opposed Nvidia buybacks a year ago, prefers this — “it keeps the thing going” — but: “the time to manage for the downside is when no one is managing for the downside… we remember that stage of the cycle in ‘99-2000. History doesn’t repeat, it rhymes.” Jason agrees nobody is managing for it: “I’ll check out of that board meeting — here’s my junior associate.”
• On Anthropic’s Samsung chip talks and DeepSeek building its own silicon: the earlier critic has softened since calling it “mad” last week, crediting two arguments from Anj Midha (likely Anjney Midha) — own the compute (“if you don’t own the keys, you don’t own the crypto asset”) and silicon optimized for your own model. The speaker still squints at app-to-model-to-chip vertical integration: “I don’t get it, but I may not be understanding the big picture.”
• The semiconductor-industry critique calls the customization rationale fake: at this volume “they’ll do you your own tape-out… if OpenAI needs a different chip, you’re going to get it.” The real driver is believing Nvidia’s margins are so high that “to survive, we have to recapture that margin” — “customized for us is just soft language” while everyone avoids being aggro and preserves relationships.

7. China’s counter-strike: Kling monetizes video, open source takes OpenRouter

• Kling raised $2.8B at $18B on ~$500M Q1 ARR, headed for a Hong Kong listing — “the most commercially successful AI video product on earth is Chinese” — just as OpenAI shut Sora down. Jason’s arbitrage question: Higgsfield (he and Harry are investors) also reports ~$500M revenue and $2M/day in credit-card billings, raising at ~$5B with Kling as just one model it runs — “is that a 3x arbitrage?” Is there a Chinese AI valuation bubble? Harry’s counter: DeepSeek raising at $50B is a gross discount to Western peers.
• Rory’s Sora autopsy: a 30-second generation costs roughly $1.30-$2 in GPU cost, so it works if you charge — Kling charges fast, Sora gave away too much — but OpenAI’s highest-and-best use of finite GPUs is elsewhere: “there’s more money in coding than consumer video.” For Kling it’s a wonderful business; for OpenAI it sat below the materiality line. Harry’s kicker: “this is why we get to invest in startups — the distractions can become very large businesses.”
• The top six models on OpenRouter today are Chinese open source, and Jason — back from two weeks behind the firewall — thinks we built that outcome: even in Hong Kong you cannot use ChatGPT or Claude. “What do you expect when you can’t access the leaders? They’re going to build something as good or better if they can — and they can. At least we know they can come close.” Jensen was right about the consequences of blocking chips and models.
• Rory’s nuance: export control may be the sober national-security call — he’s “not equipped to assess that” — but “actions have consequences”; you can’t expect China to say “okay, you caught us, we give up.” The new wrinkle Harry flagged: China may deny overseas access to its open-source models — “zany” mutual distrust (“we’re nervous about using them… they’re worried about letting us use them”). If it happens, it’s a gift to US frontier models and US open-source players like Reflection and Poolside.

8. Tokens go open source, dollars stay frontier

• Everyone is pushing toward open source on cost — the DoorDash co-founder announced a move that day — but Jason’s live counterexample cuts the other way: an algorithm he couldn’t crack in 10 hours and ~$500 on Replit’s Sonnet-plus-open-source mix fell in 20 minutes to Fable plus Opus. “It wasn’t just more expensive, it was cheaper… I’m not sure I want to waste a day on a mediocre answer that doesn’t work.” His frame: “sometimes you go to the nurse practitioner and sometimes you go to the heart specialist.”
• Rory’s synthesis, via Jesse Zhang’s (Decagon) post: frontier models for unbounded problems and unknown unknowns, open source once the answer is commoditized — which is why OpenRouter token share misleads: “all the tokens can be in one place, but all the dollars can be in the other place.”
• The CX stress test: the industry is standardizing around ~50 cents per resolution (down from a dollar), implying LLM costs of 25 cents or less — hence the open-source rush (and Fin’s $3.6B sale). Jason sees plateauing in AI-CX data and suspects cheap models cap quality; Rory reframes it as the market that already crossed the chasm — customers can articulate the ROI (30% to 65% resolution rates), and if “from 65 to 75 it’s two bucks a pop, you’ll still happily pay it” — “a high-class problem.”

9. Everyone lives long enough to become IBM

• Microsoft is deploying $2.5B and 6,000 forward-deployed engineers inside enterprise clients (Amazon moved two days earlier), aimed at the MIT finding that 95% of enterprise AI pilots deliver no measurable P&L impact. Harry’s flat call: “it’s going to fail” — not on spreadsheet logic but on talent depth. His evidence: a public-company vendor told his portfolio a bug would wait three months because the good FDE was on paternity leave. “There’s just not a couple hundred thousand people that want these jobs that are off-the-charts smart.”
• Rory disagrees — and doesn’t buy the 95% stat — because corporate America genuinely can’t adopt alone, and someone must sit between the banks and oil companies and product-to-the-core OpenAI/Anthropic. “Every technology company either goes bust or lives long enough to become next generation’s IBM”: Microsoft is now the trusted incumbent selling adoption of other people’s technology, the HP/IBM Global Services press release of 20 years ago rewritten. Caveat: “it won’t be nearly as profitable as selling operating systems.”
• The stakes, per Rory: the rate of diffusion is the single biggest question in AI. OpenAI and Anthropic reached $12B and $4B revenue faster than any technology in history — “if the next 10x takes three times longer because corporate America can’t adopt, that’s going to have consequences,” specifically how fast Anthropic gets from $4.5B to $40B to $80B. Jason’s concession: it’s a slowdown, not a stop, and even flawed FDEs beat “trying yourself, which is often hopeless.”
• The working model that exists: Harvey puts an FDE plus a lawyer on every single deployment — tech expert paired with domain expert. Rory: when the vendor sells intelligent answers about your business rather than a database, “you better be damn sure those answers are grounded in oil and gas facts” — which is also why these services businesses will be tricky to build.

10. Venture’s new order: brands go personal, employees are single-shot VCs

• Ashton Kutcher is leaving Sound Ventures — roughly a billion raised, strong OpenAI/Anthropic SPV record — to start a seed/pre-seed deep-tech firm with Morgan Beller (ex-a16z, then NFX). Jason wants the gossip (“it’s just crazy to leave your own firm”) and files it with Jack Altman raising a half-billion solo fund then joining Benchmark: moves that “make sense in 2026.” Jason sees no dark story: Kutcher is one of the few investors whose name exceeds the firm’s — “he gets more searches than Sequoia without even blinking.”
• The ElevenLabs $22B secondary matters less for price — “consistent with other rounds” — than for what it signals. Jason: “as an employee today, why would you join something that you don’t believe will have secondary options? I really think this is a big issue… Life’s too short, man.” Only a handful of companies tender like clockwork (Databricks, OpenAI); Clay managed one at $5B, but a tier below that, softness kills the program.
• The correction — worth keeping: join a company already tendering and your grant prices that in. “The whole trick is to join something that isn’t doing a tender offer today, get a healthy grant, and join a company that within a year or two, when you’ve vested 50-60%, starts doing tender offers.” Employees are “single-shot VCs” — “we get 20 shots in goal and they get one” — and tenders are the proxy for an IPO window now stretched to 12 years.
• Jason’s closing self-deprecation on VC value-add, a fitting kicker: “I’m actually not here to give you insights. I’m simply here that if we’re driving the thing off the cliff, I scream stop. That’s probably the only value-add.”