DeepSeek Raises at $50B | The Rise of Open Source vs OpenAI & Anthropic | OpenAI Builds Own Chip

Summary

• The episode’s core thesis: open source is “a bit of a fake because China is paying for all the training.” DeepSeek’s $7.4B Series A at a $50B price crystallizes it — the founder commits 20 billion yuan (~$3B, roughly 40% of the round) himself, fewer than 10 investors including JD.com get in, and none get rights: only the Chinese state retains governance control. Harry’s valuation frame: with the two US closed-source leaders trading “plus or minus a trillion,” $50B for the open-source competitor “feels roughly right” — and Z.AI is already public in China around $100B, “a thousand times revenues.”
• Number three closed-source is the kill zone. Routing across models exploded in the last 90 days, open source is ~5x cheaper, and Rory’s oligopoly rule: when an industry gets ground down, “the number one guy makes a little less money, the number two guy makes quite a lot less money, and the number three guy goes bust.” Google survives only because the balance sheet sits behind GCP-style economics — and a closed-source number four is now “almost impossible.”
• The $725B question, resolved into one number: ~7-8% of the US labor force must be replaced by tokens for the capex math to work. Harry rounds $750B annual capex up to $1T of needed revenue (electricity included), requires customers to extract more value than they spend, and lands on “a dauntingly high bar.” Meanwhile capex went from 60% to 120% of Mag-7 free cash flow, funded by borrowing — “you can be intellectually right but the narrative can keep going a long time.”
• 2027 is “show me the ROI.” Token maxing was 2025-26’s tuition (“just go build it, guys”); now CIOs will allocate tokens to divisions with proven returns, not the best PowerPoint. The margin stack is the vulnerability: free users subsidized, $200 max plans buying “$10,000 of inference a month,” and 40-70% gross-margin enterprise inference — which is exactly the segment open source attacks. Anthropic is already emailing users to cache prompts so it can price under open source.
• OpenAI’s Jalapeno chip (with Broadcom, claimed 50% cost cut) is really about the “flabby middle.” Jason’s read: the labs have Sonnet/Opus at the top and Haiku at the bottom but no affordable middle model, and that middle is where open source hollows them out — halve inference costs and open source is “probably only twice as cheap.” Harry’s rebuttal is the episode’s best line: “the whole reason the OpenAI and Anthropic models work is because other idiots have spent the $300 billion on their behalf.” Cerebras fell 16% on the news.
• Moats are dying by LLM lift. Accenture down ~40% for the year: its core SI business — five-year, floor-of-consultants Salesforce and SAP deployments — is precisely what LLMs automate, and AI-first entrants bid $15M against Accenture’s $80M. Databricks claims 30-day data migrations; “your moat can be LLM-lifted away.” Public-market simplification: everyone selling seats is getting crushed; everyone selling variably is winning. “There’s only one thing worse than a seat-based model, and that’s a model that’s based on bodies.”
• Talent and labor compound to the winners. Google loses Noam Shazeer and Nobel-winner John Jumper in 48 hours because number one can “promise everything to everyone”; Harry built an AI VP of Finance from China “in a single-digit number of hours” that’s “better than any human on the team”; and the new startup deal is small, top-paid, in-office teams working six-plus days: “you want an Omega or you want to be rich. Make your choice, boys.”

Deep dive

1. Google loses two generational scientists in 48 hours — winners buy whatever they want, including people

• The departures: Noam Shazeer — co-author of the original attention paper, who worked on Character AI and was brought back to Google in a ~$2B acquisition, “made plus or minus a billion dollars” — and John Jumper, Nobel-winning co-creator of AlphaFold who joined Anthropic having worked nowhere but DeepMind since academia. Harry notes Shazeer likely left half his package unvested — that’s how strong the pull is.
• Two dimensions, not one. Jason’s frame: elite researchers “will only do what they want to do” and labs must build that environment (his college-age son instantly turned down an Anthropic internship offer that found him, purely over the learning environment). Harry’s addition: it’s also frustration at not being able to ship — Google had a ChatGPT alternative and “the bureaucracy just kind of smothered the product.” Anthropic’s edge is doing both: research freedom and executing “tactically brilliantly.”
• Harry floats the rumor that Jumper joined because Anthropic has had a secret breakthrough. Jason’s skepticism is worth keeping: the lag from medical invention to economic value is 10-15 years — protein folding “has collected its Nobel prize and as yet has not had meaningful commercial success or a drug in production.” More likely it’s just where to spend the next five years.
• Rory’s correction to Google-is-crushing-it: only Nvidia and Google among the Mag 7 have beaten the S&P over 18 months — “Facebook, Amazon and Microsoft are doing s*, they’re not relevant here” — but “you don’t wake up every morning and say let’s try the new Google coding tool. You do try Claude Code.” Relevant, funded, and definitely number three in innovation.

2. DeepSeek’s $50B round: the only shareholder that matters has an army

• The terms are the story: $7.4B at a $50B price, the founder committing 20 billion yuan ($3B, ~40% of the round) himself, fewer than 10 investors including JD.com — and none get rights. The Chinese state retains governance control. Rory: “the only people getting voting rights are the only people who don’t need them… they have an army.” DeepSeek’s original claimed $15M training cost? “Of course it wasn’t true.”
• Jason, just back from two weeks in China: Anthropic and OpenAI intentionally block China and even Hong Kong (“it’s them blocking it, not China”) while Gemini doesn’t — “it’s DeepSeek and Gemini, those are my friends.” DeepSeek is deliberately crippled domestically: no web search, and as far as Jason can tell, trained on different data. The sovereignty logic clicked in person: China “does not want to be reliant on Anthropic and OpenAI to run the next generation economy,” and versus an aircraft carrier the subsidy is “a drop in the bucket.”
• Harry’s both-sides caveat, unprompted: the US is doing its own version — Anthropic unable to ship its most recent model until it satisfies US government concerns. Credit to the 2024 Situational Awareness essay (likely Leopold Aschenbrenner; “Leo” as spoken) for calling that governments would move in. “You can’t give China grief for this kind of thing when we’re doing the same thing ourselves.”
• On European sovereign models (the Mistral question): declaring sovereignty means “I couldn’t compete as the fourth player worldwide, but I can compete as the dominant player in Europe.” Rory: “a million years of economic theory explains why that’s a dumb idea” — everyone in Europe gets a worse model at a higher price — but a government may choose to pay that political-security tax.

3. “Open source is a bit of a fake” — and it’s aimed at number three

• Jason’s mechanism, spelled out: token budgets are real, so everyone except the smallest startups is now routing across models — a shift that “in maybe 90 days on the pod” went from unclear to universal. Slot one and two go to the leaders; the real fight is whether slot three goes to the third closed model or the best open one. Open source “in inference and training is not free, unlike Linux” — but it’s materially cheaper, because “China is paying for all the training.”
• The counterpunch arrived in Jason’s inbox the morning of recording: an Anthropic email — “your prompt cache hit rate is low” — pushing discounted cached prompts that “can actually be cheaper than open source.” Not aimed at Gemini; aimed squarely at open source.
• Rory’s structural read: tech markets settle into tight oligopolies where revenue concentrates in one and two, and when a 5x-cheaper alternative grinds the industry, “the number three guy goes bust” — Google only keeps punching because of the parent balance sheet, and a closed-source number four “is almost impossible… the market is set.”
• On GLM 5.2 beating GPT 5.5 on coding benchmarks: “all it says is these guys are cranking” — roughly six Chinese open-source models, three at or near US performance, providing “a competitive drag on what Anthropic and OpenAI can charge.” Distillation from frontier models is part of the story, but they’re there and they’re relevant.

4. AI capex is repricing everything — your iPhone, your electricity, your job

• DRAM contract prices rose 90-95% in Q1 alone; Tim Cook told The Wall Street Journal that Apple faces a “100-year flood” in memory costs; memory up 4-5x in cases. Rory’s first reaction: “profound regret on not buying SanDisk and Micron a year ago and making a 20x.”
• Their mechanism is the episode’s cleanest economics: AI investment commands resources through the price system — “it’s going to manifest itself in the price of your iPhone… your electricity… your house in San Francisco… in 20% of you losing your jobs” at all-in-capex companies like Oracle. “Economics just sends its signal — it doesn’t have morality.” Jason adds that Apple will raise prices rather than eat margin, and sell a few fewer iPhones.
• Goldman projects $7.6 trillion in cumulative AI capex 2026-2031. When David Cahn’s Sequoia piece ran, capex was ~60% of Mag-7 free cash flow; it’s now 120% and debt-funded — conviction doubled rather than wavered. Rory: “this is the classic thing about bull markets… you can be intellectually right but the narrative can keep going a long time.”

5. The $725B question: the math requires replacing 7-8% of the US labor force

• Harry’s back-of-envelope, revisited: round $750B of capex up to $1T of required revenue (you pay for electricity too); customers must get more value than they spend, call it $1.5T; total US labor spend is sub-$20T — so 7-8% of the labor force replaced by tokens for the last dollar of capex to earn a return. “A dauntingly high bar… there’s a little part of me that says I don’t know if that last dollar in capex is going to earn a return.”
• Jason won’t short the bull: demand is “an order of magnitude more than today if it could be served cost-effectively — we would all be consuming tokens 24/7 if we could,” and in venture “you don’t make money not deploying your fund.” Harry sharpens why this cycle differs from SaaS: Salesforce was binary — need it, buy the seats — whereas intelligence has infinite demand at zero price, so “price is going to be the arbiter of how much you can do.” Allocating it is a CIO skill that didn’t previously exist.
• The labs are running “an A/B test”: a subsidized free tier (bigger at OpenAI), a massively subsidized prosumer tier — $200 max plans consuming “$10,000 of inference a month,” “the whole openclaw drama” — and enterprise API customers at 40-70% gross margin on inference. That lucrative enterprise segment is precisely what open source attacks.
• The 2027 turn: token maxing was rational tuition (“it was the best way to get teams AI-fluent”), budgets then blew, and next comes “show me the f*ing ROI” — tokens to the division that laid off 20% or grew fastest, not the best pitch. The trap is the parity tax: if everyone adopts, productivity gains vanish from relative profitability (Jason’s ATM-and-tellers example) — “but the people who don’t adopt it are dead,” so everyone leans in and “our team has to be 15% leaner next year, guys.”

6. The AI VP of Finance: agents replace what humans are unwilling to do

• Built from China “in a single-digit number of hours” with “a very mediocre prompt,” after a human forgot to invoice $80,000 that had to be written off: the agent creates the quote, builds and ships the contract, updates Salesforce, invoices through bill.com, chases payment, interacts with Brex, closes the transaction in QuickBooks — “for the first time in 10 years our books are accurate.” Harry’s verdict: “it’s better than the humans.”
• His honest reframe: this isn’t replacing the team — “we really only want to do like 5 or 10% of our jobs, including investing” — agents do the parts humans won’t (follow-ups, proper invoices). But variable-cost contractor spend “could fall by 50 to 60% without even intentionally trying to save money.”
• Brandon from Record’s (as spoken) claim — “training agents will be the largest job category in five years” — meets Jason’s pushback: “didn’t we say this about prompt engineers when the show started?” $150k-out-of-college prompt engineering is “worthless today.” Harry’s synthesis: the skill is real but redefined every year — mastery means knowing where agents break. Their finance agent admitted it didn’t fully read a contract; asked why: “I don’t have a good answer for you.” “Sonnet rapidly goal-seeks” — and looping, self-improving agents will change the discipline again.

7. What VCs fund now: margins, small funds, and six-day weeks

• Nicolas Dessaigne of YC (likely; Algolia founder as referenced) tweeted that good companies now fail Series A/B on margin, not growth: “investors don’t fund revenue, they fund the margin on it.” Rory’s disagreement is categorical about the past: negative-gross-margin hypergrowth “has in fact been 100% true” as a strategy — it describes the foundation models, the inference providers, and Cursor, which grabbed the ground and is “worth 60 billion.” But he concedes investors may now be tightening for the next generation.
• Jason’s tell that the era is ending: all three of them are “sitting on a portfolio company investment we made in the last 12 to 15 months where inference was the marketing strategy” and they’re unsure they’ll “get right side up.” Without massive growth, “these startups will just fail” — and no one acquires an adjacent negative-gross-margin business in a downturn.
• On Menlo raising “only” $3B after 50 years and the Anthropic win: the smart structure is a conservatively sized main fund plus SPVs on demand for the two-or-three-a-decade trillion-dollar anomalies. Rory’s mechanics: smaller funds carry higher risk-adjusted returns (less cross-deal aggregation), and “your fund size dictates your strategy” — in Q1, ~70% of venture dollars went to four or five deals.
• The rage-bait segment lands here too: Ryan Petersen’s “work from home is white collar fraud” clip is, per Jason, simply dated — new portfolio companies want small, top-of-market-paid, double-equity teams in office six-plus days a week. What was “utterly toxic” when Cognition said it 60 weeks ago “is how you build a winner” today. “You don’t get to make 10 million for working 18 hours a week. You get a watch. You get an Omega. You want an Omega or you want to be rich. Make your choice, boys.”

8. Kalshi at $2B run rate: a regulatory arbitrage riding America’s urge to gamble

• Jason’s “end of complex analysis”: “Americans like to gamble,” the Supreme Court unblocked it, and Kalshi — 80-90% sports betting despite the prediction-market label — found CFTC federal jurisdiction that lets it skip the state-by-state licensing binding FanDuel and DraftKings. Structurally it’s a clearing house matching buyers and sellers, not a betting house, “but to a rounding error the experience is much of a muchness.” IPO talk at ~10x on $2B revenue; growth so fast it “could only be seven or eight times” by listing.
• Two risks, one opportunity. Regulatory: states are suing, and “is there a chance that in a non-Trump administration the party stops? Yes — that’s just part of the bet.” Competitive: Jason thinks Meta could “clean up” with social in-feed betting if it’s willing to cut the same corners; Rory’s counter — the sports-betting demographic is young and male, “not the Facebook demographic within a million years,” though the aging user base is exactly why Meta is looking.

9. Accenture -40%: the death of moats, bodies, and seats

• Two things at once for Accenture (down 19% on the day, ~40% on the year): helping companies adopt GenAI is exploding from zero, while the 90%+ core — SI consulting — is among the markets “most prime for disruption by AI.” Rory’s proxy: anything a company outsources to India it will outsource to AI, so read the BPO spend map — Accenture sits at the top of it. AI-first entrants now bid $15M against an $80M Accenture proposal, and incumbents can’t respond because “there’s only one thing worse than a seat-based model, and that’s a model that’s based on bodies” — bill 100 people at $500k who cost $200k, and needing only 40 collapses the margin structure.
• Harry’s specimen from Adobe: “an entire floor of Accenture for 5 years deploying Salesforce” at an estimated $26M+ a year. Now Databricks — “growing 80% at 6 billion or something like that, accelerating” — claims a 30-day LLM lift of all your data, and Salesforce moved 20VC off Marketo “in a couple weeks… with no humans” after five years of failed attempts. “Your moat can be LLM-lifted away” — a founder droning about moats in a pitch this week made Jason instantly not want to invest.
• His public-market simplification: “everyone selling seats for the most part is getting crushed; everyone selling variably one way or the other is winning” — variable pricing attaches to AI spend and a strong economy, seats attach to headcount that’s shrinking. Harry agrees on the trend but notes the near-term crush is less about seat counts than about the consulting floor being the easiest work LLMs can do.

10. Jalapeno and the “flabby middle”: why OpenAI built a chip — and why Harry thinks it’s backwards

• The announcement, dropped mid-recording: OpenAI’s inference chip co-developed with Broadcom, claimed to beat state-of-the-art GPUs on performance per watt, with Broadcom’s CEO on record that it cuts costs 50% versus a typical GPU. The market’s verdict was immediate: Cerebras — whose big traction was a $20B chip order from OpenAI — fell 16%.
• Harry’s case against: with three-plus chip vendors and five-plus hyperscalers “breaking their picks to provide you with cheap compute and taking on all the capital risk,” the largest buyer on the planet should extract a 20% discount, not vertically integrate two levels backwards. “The whole reason the OpenAI and Anthropic models work is because other idiots have spent the $300 billion on their behalf.” His cycle-top tell: “if you vertically integrate back into memory, then you know it’s over. Just you and the Koreans.”
• Jason’s case for — and the sharper strategic claim: the labs have Sonnet and Opus at the high end, Haiku at the low end, and no middle product, because the middle is too expensive to serve; that “flabby middle” is exactly where open source disrupts “just when they’re all ready to go IPO.” Cut inference costs in half and “open source is probably only twice as cheap for a lot of use cases” — the middle market becomes servable. Both agree the chip decision was made “in a different era… of abundance,” and Jason’s closing wish: Google should “wake up” and price-pressure Sonnet, Opus, and GPT 5.5 with a nearly-as-good, US-based, competitively priced middle model. “They just haven’t made it happen.”