20VC: Anthropic's $2T IPO and Instinct's Rumoured $30B Round

Summary

• Anthropic’s S-1 establishes a new benchmark for venture scale, speed and liquidity. The panel highlights $16.3 billion of first-half revenue, stronger-than-expected profitability measures and $110 billion already on the balance sheet. Jason Lemkin’s comparison is five years to IPO versus SpaceX’s 24: “Time value to liquidity still matters.”
• Anthropic’s model economics combine a short product life with an enormous capital barrier. Some 80% of revenue comes from models under six months old, while roughly 70–80% of compute spending goes toward training. Rory O’Driscoll sees both the danger of missing a generation and the protection afforded by a business requiring “billions and billions and billions of dollars just to play.”
• The AI infrastructure bill is moving into equity issuance and potentially deeper staff cuts. Meta’s proposed raise reaches $95 billion, while Microsoft has offered voluntary buyouts to 60,000 of 223,000 employees—not announced that all will leave. Rory expects depreciation to pressure operating income more visibly in 2027, making headcount the adjustment mechanism: “This is how it’ll be paid.”
• Chinese open-weight models face a trust problem that could strengthen closed-model pricing power. Jason cites Cohere’s findings about Qwen sending data to Alibaba servers, while explicitly saying there is no evidence of an intentional backdoor. A panelist asks what shrinking Chinese-model use means for inference providers; Rory argues that even 10% of this enormous market could sustain alternatives: “It’s the mother of all markets.”
• AI-driven vulnerability discovery leaves neglected software exposed faster than owners can repair it. Harry cites two critical PHP flaws hidden for more than a decade; Jason’s concern is the installed base without functioning maintenance teams. Rory expects “a year or two of this,” while Jason sees attackers’ limited ability to extract money as one possible brake.
• Instinct’s rumored $30 billion valuation exposes a sharp disagreement over evidence and upside. Harry would invest for the small chance of a Western super app, arguing that liquidation preferences and strategic buyers protect the downside. Jason wants 100 days of traction and rejects that comfort around a consumer product with negative gross margins: “I’m just tired of doing growth rounds with one month of traction.”
• Jason proposes putting 20% of the fund into OpenAI at $1.25 trillion as a valuation dislocation. He argues that OpenAI has closed competitive gaps while Anthropic has momentum. Harry points to Codex’s improved quality and teams switching from Claude Code, and says OpenAI has achieved parity. Jason’s categorical claim: “Worst case here, we double our money.” Rory finds the proposal plausible if OpenAI’s trajectory is converging and capital discipline improves, but says continued burning would change the case.
• Mega-fund deployment is accelerating against unusually narrow public-market strength. Harry reports LP complaints that platforms raised bigger funds, spent faster and will return next year; Rory says only 40–50 S&P stocks are working, largely around technology and AI. Anthropic’s success keeps capital chasing the next exceptional outcome: “We are risk-on until we’re not.”

Deep dive

1. Anthropic resets the venture benchmark

• Rory calls the S-1 “a really top-tier performance”: $16.3 billion of first-half revenue, better-than-expected margins and operating income, including after stock-based compensation, plus reassuring customer concentration and compute efficiency. Q3 revenue remains undisclosed, but his overall judgment is emphatic: “The burden of proof is most definitely on the side of the naysayers.”
• Jason’s framing is liquidity speed: five years to IPO versus SpaceX’s 24, potentially distributions inside six years versus other investments taking 18 years. “Markups are fun,” but realized returns change the comparison. Rory agrees that everything else looks like substantially more work for less return.
• The balance sheet reinforces that advantage. Rory cites $110 billion on the balance sheet after raising $130 billion and spending only $10–11 billion. Jason highlights Byron Deeter describing Anthropic as Bessemer’s greatest investment, alongside its largest checks: scale, ownership and speed arrived together.

2. Six-month products make capital access part of the moat

• Rory’s unsettling statistic: 80% of revenue comes from models less than six months old. Anthropic does not present a conventional gross-margin calculation; footnotes suggest roughly 70–80% of compute costs fund training. Serving completed models looks highly profitable, but the next product must arrive quickly. “If you miss even once, you’re screwed.”
• His counterargument is that rapid replacement can favor the leader. Against roughly $8 billion of compute spending year to date as of June 26, he floats $16 billion next year, explicitly as an estimate. Very few competitors can write $10–12 billion compute checks; the short shelf life forces challengers to keep financing the next generation.
• Jason identifies a financing warning outside the S-1: Broadcom reportedly had to guarantee most of a related debt issue so Anthropic could buy its chips, and Anthropic paid higher interest rates than expected. Jason says banks are finding it harder to sell AI-lab debt, raising the concern that debt is exhaustible. Rory accepts that banks may have reached their debt-digestion limit, but says “we’re equitizing the issue”: approximately $100 billion of Anthropic stock issuance, alongside Meta’s $70 billion and Google’s roughly $80 billion in stock and SpaceX’s IPO, can keep spending moving.
• On reception, Rory says Anthropic should do well, then qualifies: “Hell, I don’t know.” He distinguishes possible Nasdaq-index buying from immediate S&P 500 admission. Jason also cautions against taking 900% NRR literally—pilots and trials might make the underlying figure more like 300–400%—while still expecting strong demand.

3. Microsoft’s buyouts foreshadow the depreciation squeeze

• Harry describes voluntary buyout offers to 60,000 of Microsoft’s 223,000 employees, roughly 27%. Jason stresses that not everyone will accept, but interprets eligibility as a “scarlet letter.” Harry’s pushback is cultural: what happens when employees stay after being told, in effect, that management would prefer them to leave?
• Jason sees an attempted humane process with uncomfortable consequences. Employees already planning to leave benefit; Rory calls their severance a “call option.” But neither an HR meeting nor avoiding holiday layoffs makes dismissal painless. Jason would prefer another decisive cut over leaving a large group feeling unwanted.
• Rory’s accounting mechanism matters more than the announcement alone. He calls 2026 a “golden period”: semiconductor sellers recognize gains while infrastructure buyers have not yet absorbed the full depreciation charge. As that bill grows in 2027, maintaining operating income could mean pushing harder on headcount.
• Jason’s sharper example is Cloudflare’s May 7 cut of 20%, despite accelerating growth and no apparent business downside yet. His broader claim is that software companies can no longer expect stability at $100–500 million of ARR. Faster competition means “we just can’t carry folks any longer.”

4. Meta’s equity raise keeps semiconductor demand funded

• Harry breaks Meta’s proposed financing into up to $95 billion: $35 billion of common and preferred equity plus a $60 billion at-the-market program. Rory separates that decision from the chip sell-off. Meta’s underlying business produces roughly $33 billion of quarterly cash flow, he says; becoming free-cash-flow negative explains turning to equity.
• Rory says semiconductor stocks rebounded after Anthropic’s S-1 because a roughly $2 trillion company could have $200 billion to spend on chips. He cites stronger memory stocks, SK Hynix and a 10% Korean-market rise, contrasting Anthropic with a $20–30 billion CoreWeave. Jason adds a divergence: the STAR 15 China market is down 40% this year despite intense AI activity.

5. Chinese model adoption runs into a trust boundary

• Harry cites DeepSeek’s approximately $15 billion raise, described as China’s largest private round, with CATL and Tencent among the biggest investors. Jason nevertheless calls this “the beginning of the end” for Chinese open-weight models in the US, pointing to restrictions beginning across federal agencies.
• Jason says Cohere found Qwen contacting Alibaba servers in China and Singapore and uploading data despite instructions against doing so. His qualification is essential: there is no evidence of an intentional backdoor; it is “probably a training mistake.” For a CISO, he argues, unauthorized disclosure remains unacceptable regardless of intent.
• A panelist challenges the logic: downloading and running weights locally should leave limited security risk. Jason disputes that premise when a tool-using model makes network calls. Rory acknowledges sandboxing but asks how anyone can know with certainty what a model does, likening the burden to supervising an employee who might leak confidential information.
• Jason says the reported behavior occurred around once per 100,000 calls—rare in a small test, material against SaaStr’s 40,000 daily calls or enterprise volumes in the millions. His alternative is practical: “Let’s just use Haiku 5.5.” He also cites Anthropic’s restored government availability and a 40% discount.

6. Inference providers retain a market, with a harder underwriting case

• A panelist challenges last week’s willingness to buy Modal at $15.7 billion if Chinese models are now losing addressable customers. Jason agrees risk has increased. He assigns a “56.3% chance” that OpenRouter proves a moment in time, suggesting Chinese open weights could fall from “60-something percent” of its usage to 10% in six months—an amount of change he says would not be extraordinary.
• Rory’s rebuttal is market size: inference competitors together represented about $5 billion versus $100 billion for the two closed frontier providers. US alternatives, including Poolside/NVIDIA, and customers for whom Chinese models are acceptable could sustain substantial demand. Even 10% is large; whether it reaches $30 billion, he does not know.
• Jason treats Cursor’s reliance on Kimi as temporary: in his account, Grok is not yet good enough to replace it, despite the $60 billion acquisition. He also questions the economics, estimating that inference providers have to pay Kimi around 30% of gross revenue to use it at scale. “The juice may not be worth the squeeze for many people.”

7. AI exposes old software faster than owners can patch it

• Harry cites OpenAI models and Wiz finding two critical PHP flaws hidden for more than a decade. Jason’s concern is abandoned infrastructure: government and nonprofit sites may shut down rather than patch, while some maintenance teams disappeared years ago. He estimates PHP supports roughly 40% of the internet and calls the exposure “Mr. Robot live.”
• Rory expects “a year or two of this.” Jason’s possible limiting factor is attacker economics: only so many targets can pay large ransoms, and only so many systems control dams or other natural resources. Attacks might decline when the payoff disappears, even as ransomware becomes more professional.

8. Instinct divides super-app believers from the 100-day rule

• The discussion concerns a rumored $30 billion valuation. Harry notes that the revenue multiple is the same as the last round; Rory says that means the valuation has not gone up. Jason says the capital need is real, not merely vanity: a hypothetical $3 billion raise would help fund inference and competition. His question is how long sovereign funds and other investors can keep supplying equity. “It feels infinite now,” but an eventual stop would create a structural problem.
• Harry would “unwaveringly” invest for the small chance of a Western WeChat handling finance, identity and security. He sees subscription, advertising and commerce-transaction revenues. Assuming approximately $4 billion raised, he argues late liquidation preferences and potential buyers such as Microsoft protect the downside: “You’ll definitely get that back.”
• Jason rejects that protection as sufficient. He wants three to four months—“100 days”—of traction, rather than one month around a consumer product he describes as having no revenue, an unclear travel-related business model and negative gross margins. A $2 billion loss could maim their hypothetical fund; he is unconvinced liquidation preferences prevent it.
• The comparison is Fireworks at $40 billion: Jason prefers measurable, compounding growth, even while questioning durability. Rory identifies the trade-off as a competitive market with bounded upside versus an Instinct outcome potentially worth $200 billion. His framework is a three-way payoff matrix: amazing, good or unsuccessful. Harry still votes yes; Jason uses his veto.

9. OpenAI’s execution failures create Jason’s proposed dislocation trade

• Jason calls OpenAI’s governance history extraordinary self-harm: Dario leaving, the CEO’s firing and return, and nonprofit-to-profit complications. He argues those distractions left Anthropic owning the coding market for almost 12 months—effectively five or ten years in AI time. Rory says management cannot change the past and must now converge operationally.
• Jason’s proposed response is putting 20% of the fund into OpenAI at $1.25 trillion. He compares it with Anthropic’s roughly $183 billion round, when relative valuation looked dislocated. Harry adds that coding teams he works with have switched from Claude Code to Codex over three to six months, and says Codex has achieved parity. Jason’s conclusion is that OpenAI has closed the competitive gaps.
• The return case is deliberately aggressive. Jason calls both businesses future $10–20 trillion companies, citing enterprise-budget penetration rising from roughly 0.1% to 2–3%. A 20% position needs 5x, dilution aside, to return the fund. “Worst case here, we double our money.” He proposes the maximum allocation permitted without LPAC consent.
• Rory’s support is conditional: OpenAI must be converging, embracing something like Anthropic’s discipline, and ideally conducting its final private round. Continued indefinite burning would change his answer. When Harry suggests owning both OpenAI and Instinct as a hedge, Rory replies, “Please don’t ever run a hedge fund.”

10. Exceptional founders command prices before technical risk clears

• Rory separates proven executives pursuing difficult technical opportunities from established operating companies. Examples include Bob McGrew’s Arda, raising $232 million at $1.5 billion; Kevin Weil’s Warp Science, raising $215 million; and Flapping Airplanes’ technically ambitious deal. The $550 million Lagora round, reported at $8.8 billion and over $200 million ARR, presents a different pricing exercise: more operating evidence, less dependence on one person.
• Harry describes Palmer Luckey’s War Hounds as genetically modifying military dogs at a $1.5 billion valuation. Rory openly declines to manufacture conviction from military-dog market statistics. His imagined investment memo is simply: “Wow, that’s a cool idea. Let’s put in 50.”
• Jason sees a sovereignty connection with Luckey’s rugged-laptop venture, Blue Chip: military dogs reportedly come from Europe, while laptop components largely come from outside the US. Market sizes remain unclear to him, but Luckey knows the procurement channels and unmet needs. Capturing an entire niche is conceivable; whether that justifies the valuations, Jason does not know.

11. Capital keeps moving while market breadth and governance grow awkward

• Harry’s recurring LP feedback: mega-platforms raised bigger funds, deployed faster and are coming back next year. Rory calls investors “Pavlovian creatures”—Anthropic rewarded the behavior, so it will repeat until something goes wrong.
• Rory describes a peculiar public market: ten-year P/E measures at all-time highs, but forward multiples looking cheaper after exceptional earnings, including gains on private investments. Only 40–50 S&P stocks are working, primarily around technology and AI, while roughly 450 are hurting. With IPO activity waiting on trillion-dollar offerings, he sees favorable conditions for platforms deploying large private funds.
• Jason flags Zuckerberg’s lawyers, board negotiations and threat to sell his California homes and leave within a week, while admitting incomplete understanding. Rory offers a process explanation: boards need protection when approving founder compensation. On the proposed billionaire tax, he says departures could damage ordinary income-tax receipts and emphasizes control over charitable generosity: “People like to give it away, but they don’t like it to be taken away.”