AI Prices Fell 35%, Yet Anthropic’s Existing Sales Grew 9x

Chamath Palihapitiya asked in an essay on X on 16 March 2026 what would happen to markets if AI made every competitive advantage short-lived. He titled it “The Collapse Of Terminal Value – What Happens If AI Makes Every Moat Temporary?” and called it “a thought exercise”. This week’s AI price data gives one part of his premise a test.

Prices are falling. An AI token price index from Ramp put token prices at roughly 59 cents per million tokens as of the previous Friday, down roughly 35% since the end of June, Axios reported on 28 October. Ramp’s index uses corporate spending on AI from Anthropic and OpenAI. Yet The Information reported that Anthropic’s revenue from existing customers has grown ninefold over the past year, citing its IPO filing.

Short explainer · AI prices brief

A vertical news Short narrated from the published headlines and opening lines of this piece and four related FourWeekMBA pieces on AI prices and Anthropic’s costs.

Business Pill · Model routing

An evergreen Business Engineer explainer on model routing: sending each request to the cheapest model that can handle it. It explains the general idea only and does not discuss these companies.

Where This Week’s Prices Point

Four comparisons from this week show the same direction. Each uses a different unit, so they are not one market price.

• OpenAI’s own claim. OpenAI says GPT-6.1 Terra matches GPT-6 Astra on Agents’ Last Exam at one-tenth the price per token, as we reported on 26 October.
• Artificial Analysis. On its Intelligence Index v5, GPT-6.1 Sol (max) scores 55 at $1.27 per task against 58 at $10.04 for Claude Opus 5.5 (max). That is about 7.9 times cheaper for three points less, our arithmetic. The firm writes that “cost per task varies by more than 13x among the leading models”.
• Mistral’s price list. It puts DeepSeek V4.1 Flash at $0.30 per million input tokens and $1.20 per million output tokens, against $1.36 and $4.18 for its own Mistral Large 4. Artificial Analysis scores them 32 and 25.
• SAP. SAP uses more than 70 AI models and increasingly chooses the cheapest one capable of a task, according to Reuters. It expects more than €2 billion in efficiency gains over the next two years.

What Was Posted on X

Chamath Palihapitiya’s article on X, posted 16 March 2026 (UTC) and titled ‘The Collapse Of Terminal Value – What Happens If AI Makes Every Moat Temporary?’, which opens by calling itself a thought exercise.

Bindu Reddy posted that, by her account, 80% of Anthropic’s revenue comes from API consumption pricing, which she calls its soft spot, and that DeepSeek costs 20 times less than Opus and works for 70% of tasks; the figures are her claims.

What Chamath Argued

His premise is a scenario, stated in one sentence: “AI lowers the cost of disruption so dramatically, and raises the pace of innovation so relentlessly, that no company can credibly project its free cash flow beyond five years.”

He then put numbers on it. “If a business faces a 20 percent annual probability of being rendered obsolete by AI in any given year, which is not an unreasonable assumption in fast-moving sectors, its expected lifespan is roughly five years.” At a 9 percent cost of equity, he wrote, “you arrive at approximately 3.9x FCF.”

He also named the limit of his own scenario: “This scenario is likely self-defeating.” In his words, “The disruption engine disrupts itself.”

What Anthropic’s Filing Shows

Anthropic’s Form S-1 shows compute and infrastructure expense falling from 251% of quarterly revenue in the first quarter of 2025 to 48% in the second quarter of 2026, when it recorded operating income of $974 million.

The filing explains it this way: “Compute and infrastructure expenses as a percentage of revenue decreased from 251% to 48% over the same period, as revenue growth outpaced the growth of our committed compute cost base.”

The Information also posted that “Anthropic’s ability to run AI models on chips from Nvidia, Amazon, Google and AMD may be helping it achieve an 85% gross margin on serving existing models.” And SAP, the buyer switching to cheaper models, told Reuters: “The goal is healthy margins as scale grows, not an absolute reduction in AI expenditure.”

The Long-Dated Bet He Flagged

The essay’s sharpest point was about capital spending: “The companies driving AI disruption are currently committing $300 to $500 billion per year to AI infrastructure. That capital expenditure only makes sense if you believe in durable returns over seven to fifteen years.”

Anthropic’s S-1 shows that kind of commitment being made. After 30 June, it says, the company “entered into new, and amended existing, lease and commercial agreements primarily to secure additional compute and infrastructure capacity, with aggregate future minimum payment obligations of approximately $245 billion”. As we reported, that is about 1.9 times its $131.7 billion of cash and short-term investments on 30 June, by our arithmetic.

Connecting the Coverage

The price side of this story has built up over a week. On 24 October we reported that Anthropic’s token prices fell about 50% in a year, according to YipitData. Artificial Analysis’s new index, covered on 28 October, priced the leading models’ tasks, and Mistral now hosts DeepSeek beside its own models. SAP is routing work to cheaper models.

Open-weight releases kept coming the same day, from Google’s EmbeddingGemma 2 to Nvidia’s Nemotron 3.5 Transcribe. On the cost side, Anthropic’s compute bill fell to 48% of revenue. What’s new here: one place that puts the price declines next to the leader’s unit economics and Chamath’s test.

The Structural Read

As we read it, Chamath’s scenario is about how long cash flows last, and the price data speaks to one input only: how fast a given capability gets cheaper. It does not show whether any seller’s customers leave.

SAP’s answers to Reuters show the buyer’s side of the same move: it picks the cheapest capable model for each task while still expecting its total AI spending to rise.

Anthropic’s S-1 says the falling ratio came as revenue growth outpaced the growth of its committed compute cost base, and the same filing records about $245 billion of new and amended obligations after 30 June.

Three Implications

FOR MODEL BUYERS Three of the four comparisons in this piece put the cheaper option at a tenth to about a fifth of the reference price, by our arithmetic, with the score differences each source gives; Ramp’s index is down roughly 35%.

FOR MODEL SELLERS Anthropic’s S-1 shows compute at 48% of revenue in the second quarter of 2026, and The Information reports a ninefold rise in revenue from existing customers.

FOR INVESTORS WEIGHING THE ESSAY Chamath’s own essay says the scenario is likely self-defeating; the S-1 shows long-dated compute commitments still being signed.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework puts it as “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.

As we read it, this week’s price data shows which layer is becoming abundant: tokens at a given capability level. The framework was published in March 2026 and does not discuss these prices, Anthropic’s filing or Chamath’s essay.

What Is Not Established

Chamath’s essay is a scenario about equity valuation multiples, published in March 2026. It does not discuss these models or prices. The four price comparisons use different units and are not adjusted for quality beyond the scores named. The ninefold and 85% figures are The Information’s readings, and the 85% figure is framed as something that “may be helping”; we did not verify either in the S-1 text ourselves.

Ramp’s index covers corporate spending on Anthropic and OpenAI, not all AI. We did not contact Chamath Palihapitiya, Anthropic, OpenAI or SAP.

The Bottom Line

Ramp’s token price index fell roughly 35% from the end of June to 23 October, according to Axios, and three model comparisons this week put the cheaper option at a tenth to about a fifth of the price, by our arithmetic. Anthropic’s S-1 shows compute falling from 251% to 48% of revenue over the six quarters to 30 June. Separately, The Information reports that its revenue from existing customers grew ninefold over the past year.