Sovereign AI Is Now Two Different Products

In August, OpenAI cut the price of its Luna model by 80%, and within days Anthropic, xAI and Meta matched it, all chasing cheaper Chinese models and profitability ahead of possible public listings ^[1]. Analysts had already begun describing the frontier models as interchangeable cola brands, a Coca-Cola-and-Pepsi rivalry ^[2]. The strange part is what happened to the bills. Unit prices fell, and total bills rose anyway. Uber burned through an annual AI budget in months. PNC, the bank, began buying its own GPU hardware to stop paying for external tokens, and Gartner projects AI coding tools could cost companies more than their developers' salaries by 2028 ^[3]. The industry has a name for this — a "token paradox" — where cheaper intelligence simply gets consumed more, so the hidden compute costs swell ^[4]. China has already run this journey to the end. Its labs now operate what the state calls token factories, producing 1.4 trillion tokens a day, alongside token trading platforms and even token loans — AI output priced and exchanged like electricity ^[5]. There is no road back up the price curve. Morgan Stanley expects open-weight models — systems anyone can download and run — to keep pressing token prices down, and says the industry's own hedge is to treat cheap model access as a loss leader for pricier cloud storage, security and managed APIs ^[6]. That leaves the question of where the profit is. Alex Karp, Palantir's chief executive, has answered it on the record.

Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane. — Alex Karpovsky

His company's books make the case. Palantir ran a 36.5% net margin — the profit left after every cost — on $4.48 billion of 2025 revenue, with the U.S. federal government as its single largest customer at $1.85 billion ^[7]. OpenAI, selling tokens, widened its operating loss to $12.3 billion ^[2]. The contrast is not universal. Anthropic overtook OpenAI in quarterly revenue and posted its first operating profit of $559 million, even as it committed to capital spending more than a hundred times that figure ^[2][8]. Nvidia, selling the hardware both camps need, just issued its first full-year guidance for roughly 70% sales growth ^[9]. But Palantir's method is built to outlast any price war: it embeds its operational logic inside client workflows, creating switching costs no discount can undo ^[10]. And it sells to the one customer that doesn't price-shop. A prominent researcher warns that frontier systems can already run coordinated cyber attacks at machine speed, which means a military needs its own AI running at the same speed — demand that no price cut reaches ^[11]. Nowhere is the split clearer than in what OpenAI did with Britain. Stargate UK was pitched as the country's flagship sovereign-compute project; then OpenAI paused the data center, citing runaway energy costs and regulatory uncertainty, amid a broader shift toward renting compute from AWS and Oracle after financing difficulties ^[12]. The data center was set aside. The memorandum to integrate frontier AI into UK public services was kept. The build was the part pitched as making Britain independent.

AI compute is foundational to that goal — we continue to explore Stargate UK and will move forward when the right conditions, such as regulation and the cost of energy, enable long-term infrastructure investment. — OpenAI

Britain was not being sold independence; it was being sold a subscription to it, dependence wearing the word "sovereign" on the label. Two opposite things now travel under that name. India's push began not with a pitch but with a switch being flipped: a U.S. export-control order forced Anthropic to abruptly disable its models for foreign customers, disrupting Tata Consultancy Services and Infosys overnight ^[13]. Sarvam AI, the Indian lab, drew the lesson.

For AI users, it is clear that you should not confuse access with ownership, or adoption itself as an advantage. — Sarvam AI

The BRICS+ coalition — Russia, China, India, Iran and the UAE — is building its own stack of models, local data centers and indigenous chips, explicitly to decouple from Western cloud vendors ^[14]. Palantir and Nvidia, meanwhile, sell a turnkey Sovereign AI Operating System to central banks and defense ministries, into a market McKinsey projects will grow from roughly $150 billion to $600 billion by 2030 ^[15]. PNC buying its own GPUs and Delhi building its own stack are the same decision at two sizes: a customer concluding the only way to stop paying for access is to own the means of it. Under one label, two products — one a vendor's recurring revenue, the other a customer's exit. Britain got the clearest look at the difference: what was kept, and what got paused.