Latest / HotTea: The AI Morning Briefing / Anthropic files for an IPO. What does the loss figure mean?
Transcript
- Sections 00:00 What we're covering today 00:30 1. Anthropic files for an IPO with $16.3 billion in half-year revenue 01:43 2. OpenAI cuts the cost of its Terra model and targets routine computer work 02:47 3. OpenAI says its AI refuted a major computing conjecture. Other researchers now need to check it 03:44 4. UF finds that more patient data helps some AI predictions, but not every task 04:52 5. Powerlattice tests a power chiplet on an AI accelerator stand-in 05:40 Visit Hot Tea Disclosure Narration uses an AI-generated voice. Transcript This is Hot Tea for Tuesday, October twenty-seventh, twenty twenty-six. This briefing uses an A I-generated voice. Anthropic's filing gives us a closer look at its finances. We'll cover OpenAI's lower-priced Terra model, then its research claims that still need outside review. A Florida study asks what more patient data changes in clinical predictions. We also have a power-chiplet test using a stand-in for future accelerators. Anthropic files for an IPO with $16.3 billion in half-year revenue. Its October 26 filing says it plans to list Class A shares on Nasdaq under the symbol ANT. The share count and proposed price range have not been set, and the registration statement has not become effective. For the six months ended June 30, Anthropic reports sixteen point three billion dollars in revenue. A year earlier, it reported one point two billion dollars. Its operating loss was six hundred ninety-three million dollars. The much larger net loss was eighty point seven billion dollars. That includes seventy-two point four billion from changes in the fair value of financing instruments, including convertible notes. It is not a measure of cash spent running models. The filing also shows customer concentration. The five largest customers supplied about twenty-five percent of revenue in the first nine months of 2026. That covers a different period from the half-year figures. Filing the paperwork does not complete the offering or establish a final valuation. The next prospectus and later financial statements will show the terms, costs, and cash requirements investors are being asked to fund. OpenAI cuts the cost of its Terra model and targets routine computer work. OpenAI released G P T six point one Terra on October 26. It is aimed at routine work across computer applications, including reading documents and entering information. OpenAI says its standard developer prices are half those of the previous Terra and six point one Sol. On Agents' Last Exam, it reports a fifteen percentage point gain over the previous Terra at medium effort, for about one-third the cost. Those are OpenAI's measurements. They do not show what it costs to finish an employer's actual work. Someone still needs to check the result, fix mistakes, and control what the model can do. OpenAI also places Terra at its Critical cybersecurity capability threshold. That is a capability rating, not proof that a deployment is safe. The comparison that matters is completed work and corrections on the same tasks, with the same evaluation budget. A cheaper attempt only helps when the completed result holds up. OpenAI says its A I refuted a major computing conjecture. Other researchers now need to check it. Its October 26 announcement also includes mathematics and physics results. The Exponential Time Hypothesis is an assumption researchers use to set conditional limits on how fast algorithms can solve certain problems. OpenAI says it formalized the result in Lean. It named Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams to lead the external group that will examine and explain the work. This remains OpenAI's claim, not an independent confirmation by HotTea. The materials are not all available in the same way. OpenAI says the computer-science materials are public. The physics group chose to withhold its A I-produced artifacts while it prepares a write-up. Outside review is the evidence to watch next. A claimed refutation does not, by itself, solve P versus N P or establish a faster practical algorithm. UF finds that more patient data helps some A I predictions, but not every task. A University of Florida update on October 26 asks when clinical A I benefits from more kinds of patient data. The comparisons cover twenty-two thousand two hundred fifty-four hospital admissions. They use structured records, radiology reports, chest X-rays, or combinations. The university reports a twenty-one point four percent gain for diagnosis prediction against the best single-source models. It reports an eighteen point four percent gain for medication prediction. Structured records alone stayed competitive for death-risk and length-of-stay predictions. These results come from the first twenty-four hours of care in one hospital dataset. They are not measured reductions in patient harm. The university says diagnosis and medication scores remain at or below zero point six, short of what deployment requires. The update is new, but that does not mean every paper it describes was published that day. Researchers still need to test the results at other hospitals and measure errors before they can support patient care. Powerlattice tests a power chiplet on an A I accelerator stand-in. In its October 26 demonstration, the chiplet powered a silicon device that Powerlattice built to imitate the changing electrical demands of future A I accelerators. The design moves voltage regulation closer to the processor. It puts the controller, power transistors, and inductors in one chiplet. This was a company demonstration on an emulated load. It was not an independent test of a production accelerator or a working data centre. The release does not show a measured customer deployment or fleet result. Production tests need to compare useful work completed, electricity used, heat, and reliability under similar loads. Until then, this demonstration does not establish lower data-centre electricity bills. That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.