Digest: AI cracks open math problems, SK hynix sells out through 2029, and the Fed decides tonight

Quick pulse for Wednesday, October 28. Two AI labs say their models knocked down a famous computer science conjecture on the same day. SK hynix just printed a quarter that looks fake and isn’t. Robinhood missed. Microsoft wants 60,000 people to think about leaving. The Fed decides tonight, then Microsoft and Alphabet report.

AI is now producing math results faster than humans can check them#

Three announcements landed on Monday, and they all point the same way.

Axiom formalized the longest proof ever written. The Classification of Finite Simple Groups runs about 15,000 pages across more than 500 articles by about 100 mathematicians. Axiom’s engineers and mathematicians ran swarms of dozens of AI agents for 24 days. The output: about 43.1 million lines of Lean, a language where a computer checks every logical step. Along the way the agents found 34 noteworthy deviations from the published record. Fixing the biggest one, in the Unbalanced Group Theorem, produced a new theorem in group theory. Richard Lyons and Ronald Solomon, who worked on the original classification, advised on the hardest gaps.

OpenAI claimed three new results. Its internal models produced a proof of the local smoothing conjecture in every dimension. If it holds, that also settles the Kakeya and Stein restriction conjectures in all dimensions. A second result says the Exponential Time Hypothesis (ETH) is false. ETH is the widely held assumption that 3-SAT, a classic logic puzzle, needs exponential time. Hundreds of papers lean on it. The third is an existence proof for quantum field theories including QCD and pure Yang-Mills, which is one part of a Clay Millennium Prize problem.

Anthropic says Claude got to ETH independently. It published a paper and a Lean proof of a 3-SAT algorithm that runs in time 2^o(n). Anthropic says an internal Claude research model found the core algorithm on Oct 16, the same day as OpenAI’s. Anthropic also says plainly that the result is asymptotic and “says nothing about solving formulas in practice.” Your SAT solver isn’t getting faster this week.

Now the caveat, and it matters. None of this has been independently verified yet. Lean checks that the logic holds. It doesn’t check that the formal statement says what the paper claims. Anthropic’s own README makes that point. OpenAI handed its results to outside groups. Ryan Williams (MIT) and two of ETH’s original proposers, Russell Impagliazzo and Ramamohan Paturi, lead the ETH review. Nobel laureate David Gross leads the physics group, which hasn’t released the AI’s materials yet. OpenAI says it wants no academic credit. Superintelligence put it best: in math, the slow part is now the reading.

SK hynix: 104.6 trillion won in a quarter, at an 80% margin#

SK hynix reported this morning in Seoul. Revenue was 104.6T won (about $78.9B), up 328% from a year ago and 32% from Q2. Operating profit was 83.3T won, an 80% operating margin. LSEG’s consensus was 98.3T in revenue and 76.7T in operating profit, per CNBC. It’s the first time quarterly sales topped 100 trillion won. The shares rose more than 10%.

Bar chart of SK hynix quarterly revenue and operating profit: 24.4T and 11.4T won in Q3 2025, 79.3T and 60.5T in Q2 2026, 104.6T and 83.3T in Q3 2026

Source: SK hynix 3Q26 earnings release.

The forward-looking parts are bigger than the quarter:

Long-term deals. SK hynix has signed agreements with 14 major customers. Those cover more than 50% of its 2027 and 2028 capacity. Terms are typically five years.
Sold out. On the call, it said DRAM and HBM volumes through 2029 are effectively sold out, counting signed deals and ones under discussion. Customers are asking for about 2x the HBM it can make, per Seoul Economic Daily.
Prices. DRAM average selling prices rose by a high-20s percentage from Q2, per CNBC.
Shortage. It expects demand to beat supply through 2030.
Payouts. It plans about 100T won in shareholder returns this year.

China wants in too. Alibaba, Tencent and ByteDance are reportedly asking Samsung and SK hynix for server memory supply deals, some running to 2035, per Maeil Business. Those are talks, not signed contracts.

The same week, the money side of AI got creative. Bank of America, Blackstone, Citi and Morgan Stanley set up a $63B special purpose vehicle. It buys Broadcom AI chips and networking gear and leases them to Anthropic, with deliveries from 2027. Broadcom backs the residual value of the first-lien portion. D.A. Davidson’s Gil Luria upgraded Broadcom to Buy and raised his target to $400 from $350.

OpenAI’s GPT-6.1 Terra: near-flagship computer use at a fraction of the price#

OpenAI released GPT-6.1 Terra on Monday. It’s the cheap, fast model built to operate software on screen: forms, invoices, video edits. On Agents’ Last Exam, a benchmark of professional computer tasks, it scores 15 points higher than GPT-5.6 Terra at medium effort for about a third of the cost. OpenAI says it performs on par with GPT-6 Astra at a tenth of the price per token. Claude Opus 5.5 still scores higher, but OpenAI puts Terra’s cost per task about 9x lower at max effort. API pricing is $1 per million input tokens and $5 per million output. It’s live in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu.

Robinhood: record revenue, a profit miss, and crypto down 55%#

Robinhood reported Q3 last night. Revenue hit a record $1.42B, up 12%. Earnings were $0.51 a share against $0.55 expected (LSEG, per Reuters). Net income fell 14% to $476M. That includes a $67M ($0.07 a share) mark-down, mostly on its Robinhood Ventures fund. Adjusted EPS was $0.60. The stock fell about 7% after hours, per Reuters.

Horizontal bar chart of Robinhood Q3 2026 revenue by line: net interest $415M, options $399M, event contracts $173M, other $142M, equities $121M, crypto $120M, other transaction $54M

Source: Robinhood Q3 2026 results.

The mix shifted. Event contracts revenue was $173M, up more than 6x. CFO Shiv Verma called prediction markets “almost a $500-million-a-year business.” Options grew 31% to $399M. Crypto fell 55% to $120M. Expenses rose 16% to $739M, driven by credit-card loss provisions and the cost of scaling Trump Accounts and Rothera. Funded customers reached 28.7M. Platform assets were $396B, up 19%. Gold subscribers hit 5.2M. October net deposits had already passed $13B, including $6B from the Dell family’s Trump Accounts gift.

Microsoft offers buyouts to about 60,000 people, then reports tonight#

Microsoft announced a Voluntary Exit Offer on Monday. It covers staff at level 70 and below in participating teams in the US, Canada, India and the UK. About 60,000 people qualify, roughly 27% of its 223,000 employees, per CNBC. They decide in early December and leave Dec 31. This follows a voluntary retirement program in April and 4,800 layoffs in July. The backdrop: Microsoft guided to more than $50B of capex for the September quarter.

Then the calendar gets busy. The Fed announces at 2pm ET, which is 20:00 Helsinki. CME FedWatch put a hold at about 90% on Tuesday. That would leave rates at 3.75-4.00%. Chair Kevin Warsh speaks at 2:30pm ET. Microsoft and Alphabet report after the close, and Amazon on Thursday. Watch Azure growth and capex next to that buyout offer.

Bitcoin’s big buyers are underwater, and one of them stopped buying#

Bitcoin is at about $80.8K this morning, up 1.6% in 24 hours (CoinGecko). It dipped under $80K on Monday, the lowest since July. On Monday we covered the ETF cost basis. Milk Road’s new angle is who is still buying at these levels.

Per Glassnode, via Milk Road, treasury companies paid about $80.2K per coin on average. Since June, their monthly buying has run at less than 10% of the November 2024 peak. Strategy bought no bitcoin last week and spent about $205M buying back its STRC preferred stock instead. Bitmine also bought back 2 million of its own shares, per The Token Trenches, even as it kept buying ETH. The treasury trade is now buying its own stock.

The ETFs are still showing up. They took in about $110M on Monday, all through BlackRock’s IBIT. October is down only about $204M, compared with roughly $6.9B pulled in May and June. Glassnode’s levels: the True Market Mean, the average price active investors paid, sits near $77.6K. People who bought in the past 6-12 months paid about $86.7K. That’s support a few thousand dollars below and a wall of breakeven sellers above. Bitcoin also hasn’t closed below its Realized Price, which it did in the 2018 and 2022 bear markets. Kyle at Milk Road expects a dip to the mid-$70s before the bull market resumes.

Every CEO Dan Shipper says agents now filter his inbox so well that he barely reads email. So Every’s head of evals, Mike Taylor, tested what agents cite. He ran 2,484 fresh sessions with Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and GPT-6.1-Sol. Each got a question plus five to eight search results, and had to cite the two most useful.

The biggest result: his “Ultimate Guide to Prompt Engineering” was cited in 2 of 240 sessions. He rewrote only the title and subtitle to say exactly what’s inside: five principles, with a worked example. It was then cited in 240 of 240. Every’s share of citations on that question went from 35% to about 60%.

The other rules are useful if you publish anything:

Write the description as the answer, not a tagline.
Lead with what the reader gets, not credentials or the problem.
Adding “(Updated 2026)” did nothing. It went from 2% to 0%.
Agents back one post per topic from your site, so pick the winner.
Agents’ own reading picks had zero correlation with his traffic rankings.

It was a simulation, not live traffic. Taylor says testing real referrals is next.

InvestAnswers: memory, Broadcom and Tesla’s safety numbers#

James at InvestAnswers was busy on Patreon. Here’s what stood out.

He doubled down on SK hynix before the print. He added to the position ahead of earnings. He then called it “sold out through 2029,” which the company’s call backs up. His Oct 27 portfolio update had SK hynix at about 3%. Tesla is 43.1%, crypto 23.6% and SpaceX 7.1%. He also flagged the Chinese memory talks and the $63B Broadcom lease vehicle above as signs the cycle has legs.

Tesla’s FSD safety numbers. Tesla’s latest report puts FSD (Supervised) at 6,006,200 miles per major collision in North America over the past 12 months. Manual Teslas with active safety came in at 2,031,221, and without it at 845,915. Tesla’s US average estimate is 660,164. These are Tesla’s own numbers. A “major” collision means an airbag deployed. A crash counts as FSD if it was engaged within five seconds of impact. Cumulative FSD miles have passed 16.7B.

Bar chart of Tesla-reported miles per major collision: FSD Supervised 6,006,200, manual with active safety 2,031,221, manual without active safety 845,915, US average estimate 660,164

Source: Tesla FSD (Supervised) Vehicle Safety Report.

A Cerebras trade. He posted a trade alert on Cerebras (CBRS). The stock closed up about 10% on Tuesday after Oracle said it will offer Cerebras inference on Oracle Cloud from 2027.

Crypto flows. He called BlackRock the only buyer: IBIT took Monday’s whole $110M of bitcoin ETF inflows, while ether ETFs lost about $89M. On the Fed, his read is no hike this week, with oil above $100 keeping it nervous.

He covered bitcoin, the Fed, FSD and Cerebras-Oracle in his latest video.

Video: InvestAnswers, “Bitcoin + Fed Proxies, Exchange Bal, Uptober Cancelled? FSD, Cerebras x ORCL”

-Q

Sources#