OSASI.ORGCommentary · Ideas for our profession
All briefs← Back to OSASI

Paying twice: why companies want AI they can own

October 6, 2026 · OSASI Commentary

When the CEO of Palantir says customers want to know they "own the means of production", and Microsoft's CEO warns that companies pay for AI twice, the case for open-weight models is no longer only a startup argument. The show's message is that open models are now an enterprise question about costs and data sovereignty, and that an American open-weight wave led by Reflection AI is about to test it. Both of the loudest voices also have something to sell.

How a brief is made

A member reads an article or listens to a podcast and sends the snippets, key thoughts or concerns tied to a passage or a timestamp. An AI assistant, chosen for the task and the setup (Claude, ChatGPT, Codex, Gemini or another), searches the full article or transcript, pins the times, checks the quotations, and assembles a draft. The draft goes through rounds of review, and nothing is published until a person approves that exact version. The cues on each post start the player at the passage discussed. Commentary is OSASI's perspective, offered for discussion, not a summary of the source. Each brief is co-authored by a person and an AI assistant. AI assistant for this brief: Claude (Anthropic).

Source podcast · The AI Daily Brief

This episode prompted the commentary below. It is one perspective, offered for discussion.

The AI Daily Brief · October 5, 2026
Why Companies Want AI They Can Own

Listening highlights · Nathaniel Whittemore, host, on his own. Times from Spotify's transcript and chapters; the main story starts after the headlines and sponsors, at 12:40.

−26:49 to −26:17 remainingIntroductions · Nathaniel WhittemoreThree conversations about open weights: safety, national security and enterprise strategy
−14:09 to −12:35 remainingFrom DeepSeek to data sovereigntyOpen models become an enterprise question about costs and sovereignty
−12:35 to −10:42 remainingReflection AI and the AI factoryAxios' scoop: an American open model for enterprises to build their own systems on
−10:42 to −7:43 remainingWashington and NvidiaCompetition over bans, and Jensen Huang's case for open source
−7:43 to −6:54 remainingAre the Chinese models good enough?Chinese models on US servers, and why fine-tuning is the story
−6:54 to −6:10 remainingWhat the technical customers wantAlex Karp: control over compute, models, data and alpha
−6:10 to −3:40 remainingPaying for intelligence twiceNadella's Reverse Information Paradox, Microsoft's Frontier tuning, open models' record day on Vercel
−3:40 to −0:41 remainingA new refereeJay Clayton, the superintelligence task force, and what it may mean for open source
−0:41 to −0:00 remainingReasons for optimism"Open models get a lot more interesting"

Tap a time above to start the player there. If Spotify shows a card about the show first, close it and the player carries on at that time. Times are approximate transcript times; ads and episode editions may shift them. This plays the full episode, not an extracted clip.

Who is talking

Nathaniel Whittemore hosts The AI Daily Brief. This episode, published October 5, opens with headlines and sponsors; the main story starts at 12:40. KPMG, Section, Robots and Pencils and Blitzy sponsor it. Quotations come from Spotify's transcript, which is generated automatically, and the Reflection details from the Axios report the episode reads.

The show's own message

Whittemore's subject is open-weight AI in the United States, and he hears it as three conversations at once: "An AI safety conversation, a national security conversation, and an enterprise strategy conversation." The third is the one he follows. Open models first made the news as a market scare, when DeepSeek's January 2025 release took, in his telling, US$580 billion off Nvidia's market value in a day. In 2026 the talk is about what they solve "for enterprises who are worried about costs and data sovereignty."

He ends on optimism: "in a world where enterprises are looking to deal with costs and issues of data privacy and data sovereignty, open models get a lot more interesting."

An American open model, and an AI factory

The news hook is an Axios scoop. Reflection AI, backed by Nvidia, is about to release an open-weight model that Axios expects to start behind the most advanced US models and level with the best Chinese open-weight ones. Its larger aim is the "AI factory": a company takes its own data, builds on Reflection's models with its own computing power, and ends up with a customized, cheaper system of its own. Reflection is piloting the idea with Shinsegae Group in South Korea, and Axios hears that other Western open-weight models are due this month.

Whittemore adds that Reflection reached a US$25 billion valuation without a product, won a Pentagon contract in May and signed a US$6.3 billion compute deal with SpaceX. Axios confirms the SpaceX and Nebius compute deals but gives no figures.

Even the big vendors are saying it

The striking part is who is making the case. Palantir CEO Alex Karp, speaking to CNBC in July, in a clip Palantir posted: "what the technical customers want is control over their compute, their models, their data stack, and their alpha." He wants to know why, if the labs' output is so valuable, "are they charging for tokens"?

Microsoft CEO Satya Nadella calls it the Reverse Information Paradox: "You essentially pay for intelligence twice, once with money and again with something even more valuable", the knowledge you reveal to make the model useful. Microsoft is not selling open models, Whittemore points out. It is selling its own models as a base that a company can tune into something "only you control", in the words of Microsoft AI CEO Mustafa Suleyman, who reports that a version tuned for McKinsey's tasks beat GPT-5.5 on quality at a tenth of the cost.

Fortune's coverage of the same remarks adds a fair caution: both arguments are good for business. Palantir's platform promises enterprises any model without exposing their alpha, and its nine-point AI sovereignty manifesto tells them "data retention is your treasure". Microsoft sells exactly the kind of private environment Nadella recommends. That may explain some of the confidence Microsoft's supporters show in its security next to the frontier labs. It does not make the concern wrong.

Is your data safe with the frontier labs?

Neither CEO accuses the labs of breaking their own terms. Their complaint is about what happens within them. Nadella calls it "exhaust" (as Fortune reports): the prompts people write, the tools agents call, and above all the corrections people make when a model is wrong. Even under a promise not to train on your business data, that work passes through the provider's systems, and "you learn very little about what the seller is learning in return." The rights also run one way. The labs train freely on the open internet, Fortune notes, but restrict enterprises from doing the same to their models, which is why Nadella says enterprises "will demand the rights to use model outputs to fine-tune and or train their own models."

Karp puts it as a customer's questions: "Who owns the data? Are the prompts secure? Is this being transferred to you?" Neither CEO offers evidence that a lab does more than its terms allow, and both sell the alternative. Their concern does not need a broken promise to be real. The terms themselves are the issue: what they let the provider keep, and what they forbid the customer to do with what it paid for.

Chinese models on American servers

One clip deserves a closer look. Whittemore reads out a post on X saying that a great deal is now being built on Chinese open-weight models, and that it does not even have to be disclosed because it runs on US-based infrastructure. That is a point about disclosure, not a security verdict. Running a model on servers in the United States settles where the data goes. It does not settle where the model came from, and Axios reports that banks and the Pentagon still balk at highly capable Chinese models over security risk, which is the gap Reflection and others are trying to fill.

The same exchange carries the episode's most useful line for specialists. To the objection that open models are not always cheaper, the answer is that they can be fine-tuned, and Whittemore agrees: "increasingly that fine-tuning is the story." The episode keeps coming back to it: Microsoft's tuned model for McKinsey, and Nadella's demand that enterprises be free to train their own models on what they get back. Fine-tuning, training a general model further on material from one field, is how a model gets more accurate and reliable at a specialty. Law shows why it matters where expertise and confidentiality meet: lawyers preferred Harvey's custom case-law model, built on public US case law rather than client files, over GPT-4 97 percent of the time. With open weights, the tuned model stays with whoever tuned it.

What this means for OSASI

The AI factory is an enterprise idea, but the reason behind it scales down: control over whose hardware runs the model, which model it is, and where the work goes. Appraisal sits where law does: specialist judgment, and client files that must stay confidential. The questions are the same ones the big companies are asking, at a smaller size: where the client file goes, who learns from it, and what it costs. The two doors are still the options, and the OSASI Marketplace is where peers compare experience with both.

Send a thought →

A like is counted once per reader per day and stores nothing about you. A thought goes by email to the person who edits these briefs; nothing you write appears on the site unless it becomes part of an approved brief. Share opens the service you pick; this site adds no trackers.

Prefer a feed reader, Slack or Teams? Follow by RSS.