The AI Pricing Fork: What Chamath Gets Right, and What He Misses

The gap between American and Chinese AI models is real and widening. But the story of who wins and loses is more complicated than a 60x price difference suggests.

The AI industry has split in two. On one side, closed-source American models from OpenAI and Anthropic charge $10 to $68 per million tokens. On the other, open-weight Chinese models from DeepSeek and Moonshot AI charge as little as $0.18. That gap is real, but the story of who wins and who loses is more complicated than the headline numbers suggest.

The fork in numbers

Chamath Palihapitiya crystallized the debate in a July 20 tweet: "The leading AI has already forked into two options. A: Closed source American that costs $26-56 per 1MM tokens. B: Open weight Chinese that costs $0.50-1 per 1MM tokens."

His numbers are directionally right but loose. The $26-56 range only captures the most expensive American tier (GPT-5.5 Pro at $67.50 blended) and the now-deprecated Claude Opus 4.1. Mainstream American models cluster between $3 and $20. And China's Kimi K3, the most advanced open-weight model from Moonshot AI, charges $6 per million tokens, placing it squarely in the US mid-range.

The real extremes are what matter. DeepSeek V4 Flash costs $0.18 per million tokens blended. GPT-5.6 costs $11.25. Claude Opus 4.8 costs $10. That is a 55x to 60x gap on the cheapest-to-cheapest comparison, and it is widening.

Who is winning the unit economics war

Chamath's warning is that American companies spending 50x to 100x more on AI will become "financially impaired." That is a second-order concern. The first-order question is whether the American model providers themselves can survive the price gap.

The answer depends on which lab you ask.

Anthropic is the standout. According to SemiAnalysis, the company hit roughly $44 billion in annualized revenue run rate and turned profitable in its most recent quarter. Inference gross margins surged from 38% to over 70% in 2026. The company confidentially filed for an IPO on June 1. Anthropic is proof that a premium-priced American model can be a great business, even with Chinese competitors pricing at a fraction of the cost.

OpenAI is the opposite. Leaked financials obtained by Fortune showed $13.07 billion in 2025 revenue against $34 billion in total costs, producing a $20.9 billion operating loss. R&D alone consumed $19.2 billion. The trajectory is improving: in 2024 the company spent $2.37 for every dollar of revenue, and in 2025 total costs consumed $2.60 for every dollar taken in. But the absolute gap remains enormous. OpenAI's own internal projections show losses reaching $14 billion in 2026 before turning profitable in 2029.

Google is the pragmatist. On its Q1 2026 earnings call, Sundar Pichai disclosed that upgrading to Gemini 3 reduced the cost of core AI responses by more than 30%. Google Cloud revenue hit $20 billion, up 63%, with a 32.9% operating margin. The company is scaling while cutting per-unit costs, a feat OpenAI has not yet matched.

The punchline: the "American AI" narrative Chamath paints as a single story is actually three different companies with three different trajectories. Anthropic is profitable at premium pricing. Google is driving costs down aggressively. OpenAI is burning cash.

The enterprise defection is real

Chamath's second point, that customers will flee to cheaper alternatives, is already happening. The Financial Times reported on July 12 that DoorDash, Siemens, and Airbnb are among the companies now using Chinese AI models alongside their American ones. The draw is not just price. Chinese open-weight models can be self-hosted, giving enterprises full control over data and fine-tuning.

This is the structural risk for the American labs. If the best open-weight models match closed-source performance at 30x to 60x lower cost, the premium pricing of Anthropic and OpenAI becomes harder to defend. The question is not whether enterprises will adopt Chinese models. They already are. The question is how fast.

Why the hardware thesis is intact

One of the most important counterpoints comes from Bank of America. Their semis team argues that the Chinese pricing advantage is a business-model choice, not a hardware efficiency miracle. Kimi K3, a 2.8 trillion parameter model, still requires roughly 1.4 terabytes of High Bandwidth Memory per serving instance. Architectural efficiencies are real (mixture of experts sparsity, hybrid attention, native quantization), but the total silicon required per query is comparable.

BofA also makes a subtler point: open-weight models multiply memory demand. Instead of millions of users sharing a single centralized HBM pool behind a closed API, open-weight models are downloaded and self-hosted across thousands of separate enterprise, cloud, and sovereign endpoints. That expands the total number of memory sockets, which is bullish for Micron, SK Hynix, and Samsung.

The Bargo Compute Tightness Index supports this. GPU rental market tightness has actually loosened since May, falling from 63.6 to 46.7 (Balanced regime), even as token demand surged 147%.

The Jevons paradox: what Chamath misses

The most important dynamic in the data is not the price gap. It is what happens when prices fall. Token consumption has risen 147% since early May while the effective blended price per million tokens has fallen to $1.95. Open-source models now account for 54.5% of all tracked token volume.

This is the Jevons paradox in action: as AI gets cheaper, people use more of it. Total implied weekly spend on AI inference is roughly $113 million and growing, even as per-token prices collapse. The market is expanding faster than prices are falling.

That does not mean every company in the chain wins. It means the winners are the ones with the lowest cost structure and the highest volume. DeepSeek, running on Huawei Ascend chips with lower Chinese labor and electricity costs, claims 70% to 80% gross margins even at $0.54 per million tokens. That is a better margin profile than OpenAI at $11.25.

The regulation wildcard

Chamath's real warning is about a potential US government mandate that American companies use only American AI models. He argues this would be "terribly self-defeating," forcing American companies to spend 50x to 100x more than foreign competitors. If such a mandate materialized, it would create exactly the financial impairment he describes.

This is currently a policy tail risk, not a base case. But the direction of travel matters. The Biden administration's AI diffusion rules already restrict chip exports to China. The Trump administration has signaled continued hawkishness on Chinese technology. A future administration could extend that logic to mandating domestic AI consumption.

For now, the market is pricing in open access. The FT's reporting shows American companies freely adopting Chinese models. The day that changes, the thesis flips.

What to watch

• Anthropic's S-1 filing. The confidential IPO was filed June 1. The public filing will reveal whether the 70% inference margins are sustainable and how much pricing pressure the company expects.
• The next DeepSeek price cut. V4 already had a permanent 75% price reduction in May. If V5 launches at even lower prices, the fork widens.
• Enterprise adoption data. The FT identified three major companies using Chinese models. If that number grows to dozens, the pricing umbrella over American labs collapses.
• OpenAI's IPO timing. The company has reportedly pushed its own IPO to 2027. If it cannot demonstrate a path to profitability before then, the public market reception could be brutal.

More research at bargo.ai/research.

Sources: Chamath Palihapitiya — original tweet; SemiAnalysis — Anthropic IPO financials; Fortune — OpenAI leaked financials; Bank of America — semis note on Chinese LLMs; Financial Times — enterprise adoption of Chinese AI; OpenAI API pricing; Anthropic pricing; DeepSeek pricing; Google Q1 2026 transcript; Bargo compute economics and token demand data (proprietary, as of Jul 20, 2026).