OpenAI teaches ChatGPT to build the pitchbook
Edition of 2026-09-11
# AI For All ### Friday, September 11, 2026
As of today, an investment bank can hand the pitchbook to ChatGPT, and it has to decide what its junior bankers are for.
The pitchbook, minus the apprenticeship
OpenAI launched ChatGPT for Financial Services on Thursday, a version of its business product ChatGPT Work built with Morgan Stanley and Evercore as design partners. It runs on GPT-6 Astra and is sold to firms rather than individuals, starting with investment banking and equity research.
The change is not the model. It is the data underneath. LSEG, Daloopa and PitchBook now feed it financial statements and earnings transcripts, and the tool can reach into a firm's existing data subscriptions. In a briefing, OpenAI's product vice president screened a takeover target, pulled peer prices into a spreadsheet and formatted the deck to a bank's own style guide. Asked about junior bankers, he compared the tool to the spreadsheet, and described analysts and bankers working 100-hour weeks.
Two things to hold onto. OpenAI will not name the banks that have signed on, and Anthropic shipped a comparable product for Wall Street a year ago. If you run a team of analysts, your training pipeline just became a product decision.
What Anthropic caught was a billing problem
Anthropic's September threat report, published Thursday, covers misuse it disrupted between December and August across seven harm areas. The section to read is not the jailbreaks. It is the distillation.
Anthropic says Alibaba, Moonshot AI and DeepSeek built pipelines to extract Claude's reasoning and train their own models on it. In two of those cases the vendor served Claude to its own users and kept the exchanges. Moonshot, the maker of Kimi, forwarded customer requests to Claude and showed the answers as its own work. Anthropic puts the Alibaba campaign at over 151 million exchanges between May and July, and the Moonshot campaign at over 23 million.
The practical reading is uncomfortable. If your staff use a Chinese assistant, the model answering them may not be the one on the label, and their questions may have become someone else's training data. Treat that as a vendor-risk question rather than a research curiosity. It is also one competitor describing another, published by the lab that lost the traffic.
Washington stops buying seats
The General Services Administration signed a new OneGov agreement with OpenAI on Thursday: 50% off usage of ChatGPT models, billed by consumption rather than by seat, with no platform fee, no minimum order and no spend commitment, inside a 27-month offer.
Eligibility runs from federal agencies down to tribal governments, reachable directly, through resellers or through cloud marketplaces. GSA says OneGov has produced approximately $1.68 billion in cost savings, most of it from AI agreements, and that this next phase moves the government from licences to consumption as the earlier offers expire.
Watch the pricing model rather than the discount. When the largest buyer in the world accepts paying per unit consumed with no minimum, it tells every vendor what volume is worth. The catch for the buyer is the one every cloud bill carries: consumption pricing only saves money if usage stays disciplined.
Google buys the power, not just the chips
Google will invest at least €13 billion in Finnish digital infrastructure, building data centres in Kajaani, Muhos and Vaala and expanding its existing site in Hamina. The part worth your attention is the energy contract underneath: a 22-year life extension power purchase agreement with the Finnish utility Fortum, tied to the Loviisa nuclear plant.
Google puts the construction phase at more than 37,000 jobs nationwide. The announcement is two days old rather than fresh, and none of the capacity is sold to anyone else.
For a founder buying compute in Europe, the signal is about supply, not about Google. The cheapest clean power in the Nordics is being signed away in decades-long blocks to the largest buyers, leaving everyone else to compete for what remains.
Weak signal
Cognition released a coding model, SWE-2, and says it lands close to the frontier of coding while costing far less than the model it measures itself against. It is of no use to anyone outside software, and the benchmark behind the claim is its own. It still measures something buyers should watch: the price of a unit of coding capability keeps falling, and that is leverage at renewal time.
Cognition, September tenth.
That is the pattern to carry into next week. The tools get cheaper, the data deals get deeper, and the electricity is signed away before the product ships.