The Anthropic Moment, Nvidia's Rebuttal, and the 60% Problem

The Threads

The Anthropic moment, and the Nvidia rebuttal

This was the week Anthropic priced above OpenAI in the secondary markets for the first time — and it was also the week every major host had to reckon with what that means. On All-In, the Core Four walked through OpenAI’s leaked Dresser memo, which attacked Anthropic’s $30B run rate as ‘inflated’ by revenue share and framed the company as ‘fear, restriction, and the idea that a small group of elites should control AI.’ On 20VC, Anj Midha told the story from the inside: in 2021, Midha introduced the founding team to 22 venture capitalists, and 21 said no. The most common question was ‘what’s GPT-3?’ The seed only closed because Amazon, not a VC, put $4B of compute-for-equity behind it.

Jensen Huang, interviewed by Dwarkesh, gave the rebuttal to the ASIC-revolt story the headlines tell. ‘Anthropic is a unique instance and not a trend. Without Anthropic, why would there be any TPU growth at all? It’s 100% Anthropic. Trainium growth? 100% Anthropic.’ The corollary — Jensen admitted his one real strategic mistake was failing to recognise that Anthropic could not have been funded by conventional VCs: ‘a VC would never put five, ten billion into an AI lab.’ Hence the reported $30B Nvidia has now put into OpenAI, the $10B into Anthropic. He will not make the same mistake at the next lab.

And Chamath, talking about Groq’s $13B Nvidia licensing deal, filled in the supplier-as-investor pattern from a third angle: Groq’s seven years of darkness before product-market fit were subsidised the old-fashioned way — a single angel writing a $10m check — but the Nvidia deal itself materialised in weeks, once NVLink Fusion created the mechanical interface. Supplier-as-investor is now the default funding model for frontier AI.

The 60% problem and the SaaS death spiral

Jason Lemkin made this week’s most pointed operational argument, and repeated a version of it across two shows. The thesis: every public SaaS incumbent is shipping ‘fine’ in-product agents that are roughly 60% as good as standalone tools like Claude Code, Replit, and Lovable. Customers will use that 60% — but they will not pay for it. Without priced-separately agent revenue, there’s no reacceleration. Without reacceleration, ServiceNow, Salesforce, HubSpot and the rest drift into the deep-value bucket at 8-11x forward P/E, on an IBM-shaped trajectory.

Bill McDermott, who runs ServiceNow, used No Priors to push back — thoughtfully. His counter-thesis: replacing an enterprise platform with LLM-only tooling costs 10x more once you total platform replacement + human capital redeploy + GPU/token consumption. More importantly, ‘people forgive humans for mistakes but they never forgive software.’ The enterprise error tolerance is structurally smaller than the consumer’s. Language models suggest next steps in milliseconds; platforms close the case. Both are necessary.

Who’s right? Lemkin has the better short-term public-markets call; McDermott has the better long-term platform case. The reconciliation is in the timing: you can’t charge for a 60% agent today, and SaaS stocks are priced for that gap to close slowly. But the platforms that bolt serious workflow depth onto LLM front-ends within 12-18 months will separate from the ones that can’t. The ratio of those populations is currently unknowable.

Back-to-the-future venture: Arthur Rock, re-drawn

Four shows this week described the same new-old venture model in their own language. Scott Nolan on Invest Like the Best traced the template explicitly back to Arthur Rock at Intel, Bob Swanson at Genentech, and Mike Markkula at Apple — investors who co-founded, ran all-hands, wrote the stock incentive plans, and lived at the office. His own General Matter is a deliberate return to that mode, attacking a cost-plus industry (US civilian uranium enrichment — we have zero capacity) where the incumbent approach has failed.

Anj Midha reached the same conclusion from the AI side: ‘we’re entering a back-to-the-future era of venture capital.’ His AMP is a public-benefit holdings company that co-incubates companies like Periodic Labs from the ground up, with Midha working three days a week in the Periodic office and AMP’s compute team one floor above. Chamath’s own heuristic — ‘when the team fractures, back the technical person the company wouldn’t exist without’ — is the same pattern expressed as a governance rule. Even Jake Paul and Jeffrey at Antifund describe their edge as co-founding Better rather than check-writing it.

The shared feature: deep operational partnership over diligence-and-follow-on. The reason: classical-era VCs won’t and can’t fund the capex-heavy, multi-year, infrastructure-weighted projects of the AI and physical-world frontier. ‘Cost-plus oligopolistic industries are prime for this,’ Nolan said, and then listed them: space launch, defense, infrastructure, civilian nuclear fuel. Four of them already have their Arthur Rocks.

Attention is the new capital — and so is energy

Jake Paul framed it bluntly on 20VC: ‘attention is more valuable than capital.’ As coding and financial analysis get commoditised by AI, what remains scarce is taste, distribution, culture — the creator-economy primitives. Antifund’s pitch is that the mass-channel distribution of its partners can move markets the way a traditional fund’s capital used to. The ambition is $10-20B AUM and a public-markets fund. Cynics will laugh; the same cynics missed Airbnb.

In a different register, Scott Nolan, Jensen Huang, and Blackstone all said the same thing about energy. Nolan’s key chart: energy-consumption-per-capita versus GDP-per-capita with R² > 0.8. The US grid has been flat since the 1990s while China is now at 3x US total production. Jensen’s answer to every supply-chain bottleneck is ‘2-3 year problem’ — except energy, which is the only permanent constraint because of permitting. Blackstone’s John Gray calls the $700B of AI capex ‘the main thing’ driving the near-term economy and the 1990s-style productivity boom to come.

Energy is the new capital at the societal scale; attention is the new capital at the personal scale. Both are finite. Both are being priced up aggressively.

Founder mode and the relentless application of force

Two shows — Lenny’s × Keith Rabois and Uncapped × Max Mullen — arrived at the same doctrine from opposite directions. Rabois: ‘The better a company performs, the more complacency creeps in. The single highest-leverage role of the CEO is offsetting it. Criticise when winning; support when struggling.’ He named Mike Moritz’s ‘relentless application of force’ as the single common denominator of the best CEOs ever.

Mullen told the same doctrine as a story. When Amazon bought Whole Foods in 2017, Instacart was seconds from an existential narrative death. Instead of retreating, Mullen declared wartime, and in 18 months signed Costco, Kroger, and the rest of the retailer holdouts. Earlier, Ravi Gupta’s $25k-a-month-of-Blue-Bottle-versus-5000-customers moment galvanised the whole company around the P&L. Both are versions of the same idea: a great company uses bad news to compress the next decade into six months.

The third version: Chamath’s ‘org charts are vestigial bureaucracy’ manifesto. You back the technical person the company wouldn’t exist without. Jensen’s 60-80 direct reports. Elon’s five-company simultaneity. All of these are the same refusal to settle, dressed in different metaphors. And the counter-intuitive kicker Rabois offers: when talented people coast, morale goes down. Great operators want to build. Deny them the opportunity, and you lose them.

This Week’s Episodes

{/* Episodes are listed in the frontmatter; the site renders them below. */}

Issue 01 is live. Twelve podcasts, 14 hours of audio, ~30 minutes to read. Next week: same shows, new threads.