OpenAI is reported to be buying Google's chips, SK Hynix has sold more than half of its 2027-28 memory capacity, and 3 more AI moves
OpenAI is reported to be buying Google's TPUs, SK Hynix has pre-sold more than half of its 2027-28 memory capacity while posting a record quarter, Meta's chip suppliers are taking $4 billion of the equity funding its AI build-out, Axiom Math published a 43.1-million-line Lean proof of the classification of finite simple groups, and Malaysian customs seized $12 million of smuggled AI chips.
This issue covers October 27 at 12:00 UTC through October 28 at 12:00 UTC.
Five developments cleared the bar in that window.
OpenAI is reported to be planning to buy Google's TPUs, the chips Google built for itself, rented to OpenAI last year, and only began selling outright in 2026.
SK Hynix said long-term agreements already cover more than half of its 2027 and 2028 memory production capacity, and that the shortage runs to 2030.
Meta's two main chip suppliers said they will take $4 billion of the equity Meta is selling to pay for AI infrastructure.
Axiom Math released what it calls the first complete computer-checked proof of the classification of finite simple groups: 43.1 million lines of Lean written in 24 days.
And Malaysian customs said it seized about $12 million of advanced AI chips carried in servers declared as electronic components.
In brief
- Chips: OpenAI plans to buy Google's tensor processing units, The Information reported on Tuesday, citing one person familiar with the matter. No price, volume or delivery date was reported, and both companies declined to comment. OpenAI is one of Nvidia's largest customers for graphics processors, and the purchase would put a rival's silicon inside the company that Nvidia has been financing. 12
- Memory: SK Hynix reported a third quarter in which revenue passed 100 trillion won for the first time, at 104.6 trillion won, and operating profit reached 83.3 trillion won on an 80% margin. The company said long-term agreements with 14 customers, including global cloud providers, already cover more than half of its 2027 and 2028 production capacity, and that demand will exceed supply through 2030. 34
- Funding: Meta's upsized $25 billion common stock sale drew about five times the shares on offer, part of a plan to raise roughly $98 billion for AI spending. On Tuesday Nvidia said it will take $3 billion of the offering and AMD said it will take $1 billion, both alongside expanded supply agreements. 567
- Research: Axiom Math released a Lean formalisation of the classification of finite simple groups, the roughly 15,000-page proof known as the enormous theorem, at about 43.1 million lines and 2.05 million declarations. The company says engineers and mathematicians directed swarms of AI agents for 24 days, and that the work turned up 34 notable deviations from the published record. 89
- Regulation: Malaysian customs said it seized about 49.5 million ringgit, or $12 million, of advanced AI chips this month, including 30 server units found at Kuala Lumpur International Airport's free-trade zone on October 15. The servers were declared as electronic components and were being routed through Malaysia to reach a destination that required a trade permit. 10
Five developments
1. OpenAI is reported to be buying Google's TPUs, the chips it was renting a year ago
The Information reported on Tuesday that OpenAI plans to buy Google's tensor processing units, citing one person familiar with the matter. Reuters carried the story at 01:24 UTC on Wednesday. 1
A TPU is a chip Google designed for its own use. Google began building them more than a decade ago to speed up search, found they were unusually good at the arithmetic behind machine learning, and they became the main alternative to Nvidia's graphics processors. Google started selling them directly to some customers this year. 2
The report gave no price, no chip count and no delivery date. Bloomberg, citing a person familiar, said final terms of any deal are not known. Google did not respond to a Reuters request for comment outside regular hours, and OpenAI declined to comment. 12
The step changes the degree of an existing relationship. Reuters reported in June 2025 that OpenAI had begun renting Google's chips; buying them would put the same hardware on OpenAI's own books. 1
OpenAI has its own silicon programme as well. It unveiled a first custom chip in June, code-named Jalapeño, designed with Broadcom, and said it planned to deploy the processor by the end of this year. The internal chip team is run by Richard Ho, who worked on Google's TPUs before he joined OpenAI. 12
Google's side of the trade is already being counted. Alphabet's finance chief, Anat Ashkenazi, has said the company expects to recognise a small percentage of revenue from TPU sales agreements by the end of this year, with the vast majority arriving in 2027. Alphabet reports its quarterly results on Wednesday, after this window closes. 1
The TPU is becoming a product line rather than an internal asset. Oracle said on Tuesday that TPUs would be available through its cloud infrastructure, which puts Google's chips on another company's price list. 2
The two companies are now suppliers to each other in different markets. OpenAI said last week that it would make its frontier models available through Google Cloud. 1
The sums behind that decision are the reason one sourcing choice matters. Amazon, Microsoft and Google parent Alphabet plan to spend as much as $600 billion combined on AI infrastructure this year, and investors are pressing them on when the spending turns into returns. 1
Three things are unknown, and each one decides how big the story turns out to be. The volume and price are unreported, so this could be a trial rack or a multi-year commitment. Whether OpenAI is buying chips or buying capacity built on them changes who carries the depreciation. And Nvidia has said nothing about a large customer buying a competitor's silicon. 2
2. SK Hynix has already sold more than half of its 2027 and 2028 memory capacity
SK Hynix reported its third-quarter results in Seoul on Wednesday. Revenue was 104.6155 trillion won, or about $78 billion, the first quarter in the company's history above 100 trillion won. Operating profit was 83.3338 trillion won at an 80% operating margin, and net profit was 39.7756 trillion won. 34
The comparison that makes those numbers legible is the year before. In the third quarter of 2025 the company reported 24.4489 trillion won of revenue and 11.3834 trillion won of operating profit, so revenue rose 328% and operating profit rose 632%. Against the second quarter of 2026, revenue rose 32% and operating profit 38%. 3
What SK Hynix sells into this boom is high-bandwidth memory: stacks of DRAM wired together and packaged next to an AI accelerator so the processor can read from them quickly. HBM4 shipments began in earnest in the quarter, and the company expects HBM4 to pass HBM3E in both shipment volume and revenue in the fourth quarter, which would make it the mainstream part. It plans to run its M15X plant at maximum capacity by the end of the year to get there. 3
The disclosure a buyer has to plan around is the forward selling. SK Hynix has long-term agreements with 14 major customers, including global cloud service providers, and those agreements cover more than 50% of its total production capacity for 2027 and 2028. The contracts typically run five years and carry a mechanism that lets prices move within a band agreed with each customer, which gives the company stable profit without fixing a single price. 34
The company's own explanation for the shortage is a supply problem, stated plainly in its release: leading-edge processes are technically difficult, new production facilities take a long time to build, and supply is not keeping pace with the market. On that basis it expects demand to exceed supply through 2030. 3
The balance sheet has changed shape with the prices. Cash and equivalents stood at 114.5 trillion won at the end of September, up 26.5 trillion won in three months, while borrowings fell 2.8 trillion won to 15.8 trillion won, leaving net cash of 98.7 trillion won. 3
Most of the profit is going back out. SK Hynix plans about 100 trillion won of shareholder returns for 2026: at least 30 trillion won in cash dividends and at least 60 trillion won of buybacks and cancellations, including the 40 trillion won announced in August and completed last week. The board approved a third-quarter dividend of 14 trillion won and a further 20 trillion won of buybacks. The dividend is 20,000 won a share, half of it ordinary and half a special dividend sharing the gains on the company's Kioxia stake. 3
The company also called off the planned listing of Solidigm, its NAND subsidiary. Chief executive Kwak Noh-Jung said the market conditions made it the wrong moment to raise that capital. 3
Somebody pays for all of this, and the company named them. On the earnings call, finance chief Kim Woo-hyun said shipments of PCs and mobile devices are expected to fall because of the cost of higher memory prices and the shortages themselves. 4
Two numbers stay hidden. The agreements are described only as covering more than half of capacity, so there is no figure for how much of 2027 is committed or at what price. And the price-band mechanism means reported spot prices for memory and the prices SK Hynix actually receives can move apart without either side saying by how much. 3
3. Meta's chip suppliers are buying $4 billion of the stock Meta is selling to pay for their chips
Meta is seeking to raise roughly $98 billion for AI spending, and the common stock portion of that ended up about five times oversubscribed. More than half of the shares went to long-only and existing investors, and the 15 largest investors in the sale took about 70% of them. 5
The pricing happened on Monday and the numbers are specific. Meta sold 40 million Class A shares at $625 each, raising $25 billion against the $22.5 billion planned at launch, and sold 250 million depository shares at $50 each, or $12.5 billion, for an underwritten total of $37.5 billion. Each depository share is a twentieth of a share of 6.50% Series A mandatory convertible preferred stock. 11
The common stock was sold through forward sale agreements, so Meta has not collected the cash. It expects about $24.9 billion of net proceeds when the forwards settle by November 19, 2027, or $28.7 billion if underwriters take their full option. 11
The new part on Tuesday is who is buying. Nvidia said it will invest $3 billion in the offering, and AMD said it will invest $1 billion. When Meta announced the raise, it said its main chip suppliers, Nvidia, AMD and Broadcom, had expressed interest in buying up to $5 billion of equity in aggregate. 67
The equity comes attached to supply agreements, which is the part that changes a data-centre plan. Nvidia and Meta said they would expand their work on Meta's Muse family of models to improve inference and training performance, with Meta deploying more Vera Rubin systems and more Vera CPUs. Jensen Huang framed it around agents: "AI agents represent a new computing paradigm — software that understands our goals, reasons, uses tools and gets work done on our behalf." 6
AMD's side is further along in one respect and earlier in another. Initial Helios rack-scale systems are installed at Meta and being tested ahead of a broader deployment, while AMD's EPYC processors already run the dedicated computing environments behind Meta's Muse agent. Meta says it has deployed millions of EPYC processors across several generations and is a lead customer for the sixth-generation part. 7
AMD also gave Meta day-one support for two Meta models on its hardware, Muse Spark Open on Instinct GPUs and Muse Glimmer on Ryzen AI PCs and Radeon graphics. Lisa Su described the pair as building "at gigawatt scale", and Mark Zuckerberg said AMD "plays an important role" in Meta's computing build-out. 7
The market gave the deal a verdict on Tuesday. Meta closed at $620.19, below the $625 the suppliers agreed to pay, while AMD rose 5.45% and Nvidia 0.90%. 11
Read as a transaction, $4 billion of supplier money against a $98 billion raise is small. Read as an incentive, it is the point: a supplier that owns a slice of its customer's equity has a reason for that customer's capacity plans to succeed, and Meta has a reason to keep buying from the companies that just funded it. 6
Two figures remain open. The rest of the $98 billion is planned to come from an at-the-market programme that has not begun, and whether Broadcom joins at the size it indicated is unreported. 5
4. Axiom says AI agents formalised the longest proof in mathematics in 24 days
Axiom Math announced on Tuesday that it had released a complete formalisation in Lean of the classification of finite simple groups, with a project page, a public repository and a blog post describing the work. The blog post is dated October 26 on the company's site; the announcement went out on X on Tuesday afternoon. 812
The classification is the list of building blocks that every finite group decomposes into. It says every finite simple group is cyclic of prime order, alternating, one of the families of Lie type built over finite fields, or one of 26 sporadic groups. The original proof runs to roughly 15,000 pages in more than 500 articles by about 100 mathematicians, which is why it is nicknamed the enormous theorem. 813
Lean is a proof assistant: a programming language in which definitions, statements and proofs are written precisely enough that a program checks every step against fixed rules. A proof it accepts has no unproved step left, provided the statement being checked is the one the author meant. 8
What Axiom released is large in a way that is easy to state and hard to picture. The project page counts about 43.1 million lines of Lean and 2.05 million declarations, drawing on a corpus of 715 papers and books. The repository is Apache-licensed, describes the development as sorry-free, and says it uses only three standard axioms. 91314
The 24 days are documented in the blog post. Axiom opened a Slack channel on October 1, split the work into 108 milestones and about 380 tasks, and started its first swarm of agents on October 2. Agents claimed milestones through GitHub issues and reported finished work back into Slack, and some agents delegated missing prerequisites to other agents. 813
The finding that will outlast the announcement is what formalising turned up in the literature. Axiom says the work exposed 34 noteworthy deviations from the published record, and that most were minor: typographical errors, errors in tables, dropped hypotheses. 8
The substantial one concerns the Unbalanced Group Theorem. Formalising an existing argument, the team found counterexamples to a closure assertion that the published route relied on, and had to build an alternative. Solomon's 1976 theory of maximal 2-components supplied the ideas, and the replacement argument rests on a new preservation-of-simple-tops theorem. Axiom published that mathematics as a 35-page paper separate from the formalisation. 8915
People who wrote the original work were consulted. Richard Lyons and Ronald Solomon advised the project, Lyons helped clarify which proof routes remained available, and both sit on the advisory board that will curate the errata list. 8
Axiom's own account is candid about what Lean does and does not establish. Lean checks the deductions that lead to the final theorem; a human has to inspect the formal statement and the definitions it uses to confirm that the statement expresses the classification mathematicians mean. The blog states that division of labour directly, and notes that a formalisation may reach the same conclusion by a route different from the historical one. 8
The engineering report is the reusable part. Build times became the bottleneck as agents produced proofs faster than Lean could recheck them, so the team distributed cached build results across workers, ran independent checks in parallel, and pushed expensive finite computations into separately checkable certificates. Early milestones also concealed prerequisites the agents had assumed rather than proved, and the team deleted tens of thousands of lines before restarting on a firmer plan. 813
What stands today is narrower than the headline. A machine has checked that the proof follows from the statements in the repository. Whether each formal statement is the theorem as group theorists understand it, and whether every entry on the errata list is right, are questions for people who know the subject, and no independent group theorist had published a verdict as this window closed. 914
5. Malaysia seized $12 million of AI chips declared as electronic components
Malaysian customs said on Wednesday that it had seized about 49.5 million ringgit, or $12 million, of advanced AI chips this month, as investigators close in on a syndicate moving high-performance semiconductors through the country's main airport. 10
The specific find came on October 15, when authorities inspecting Kuala Lumpur International Airport's free-trade zone found 30 server units containing advanced AI chips. Aspalila Ag Tuah, the airport's customs director, gave the details at a press conference on Wednesday. 10
The method was paperwork. The servers were declared as electronic components, and Aspalila said the shipment was routed through Malaysia so that no restrictions would be imposed while the goods were exported to the next destination. The consignment originated in one Asian country and was to be re-exported to another in the region, a move that required a trade permit the shipment did not have. 10
A Malaysian company that facilitated the shipment has been called in to assist the investigation. Customs linked the consignment to a network behind a similar attempt foiled in June, and said at least five cases involve the syndicate, with seizures worth more than 100 million ringgit in total. 10
The legal basis for the seizure is a control Malaysia imposed last year on moving advanced AI chips of US origin, which it brought in after pressure from Washington to stop the flow into China. 10
For anyone modelling supply, this is the enforcement side of the shortage described in the second item above. Controls on where chips may go create a price for getting around them, and a resale market exists because the servers are worth more at the destination than at the origin. 10
Customs declined to give more detail, citing an ongoing investigation, and no charges have been announced. The origin country, the destination country and the identity of the Malaysian firm are all still unreported. 10
The read-through
Four of today's five are the same problem seen from four positions, and the problem is that committed capacity has become the scarce good. That reading is mine, drawn from the documents below; none of the companies framed it that way.
SK Hynix is the clearest case. It has sold more than half of what it will make in 2027 and 2028 to 14 customers before building it, and its chief executive described the effect in his own words: "our business structure is transforming into one with earnings stability and mid-to-long term demand visibility". 3
Meta is the second position. Its raise is large, and the detail that carries information is that the companies selling it the accelerators are putting $4 billion into the stock sale that pays for them. That is a supplier underwriting its customer's balance sheet, and it ties a capacity plan to an equity holding. 67
OpenAI is the third. The largest single buyer of Nvidia accelerators is reported to be shopping at the company that builds the main alternative, days after agreeing to distribute its models through the same company's cloud. When capacity is the constraint, vendor loyalty becomes a line item rather than a strategy. 1
Malaysia is the fourth, and the least comfortable one. The same scarcity that lets SK Hynix sell two years forward and lets Meta name its price also makes a container of servers worth mislabelling, and customs officers at one airport are now part of the supply chain's cost structure. 10
Axiom is the exception, and the reason it does not fit is worth stating. Its output was checked by a machine anyone can run, so the result does not have to be allocated by contract or defended by a supplier's equity stake; it either type-checks or it does not. The 43.1 million lines are the one thing in this issue a reader can verify tonight without asking anyone. 14
If that reading holds, the number to watch is not the price of anything. It is the share of 2027 and 2028 capacity already spoken for across the memory makers, because that figure sets how much of the next two years of AI build-out is still open to a buyer who has not signed yet. SK Hynix has told us its own number. Nobody else has. 3
What would change today's picture
- OpenAI or Google confirming the chip purchase, with a volume, a price or a delivery date; a denial would settle it the other way.
- Alphabet's results call on Wednesday giving a figure for TPU revenue recognised or contracted, which is the first hard number for how big that business is.
- Nvidia commenting on a large customer buying a competitor's silicon, and whether any agreement between the two companies changes.
- SK Hynix disclosing what share of 2027 capacity specifically is committed, or any of the 14 customers disclosing what it agreed to buy and at what price band.
- Meta starting the at-the-market programme that makes up the rest of the $98 billion, which fixes the real cost of the raise.
- Broadcom saying whether it took equity at the size it indicated.
- A group theorist unaffiliated with Axiom reviewing the formal statement of the classification, or the errata list, and publishing a verdict either way.
- Malaysian authorities naming the origin and destination countries, or a second country reporting a seizure linked to the same syndicate.
References
- 1
- 2
- 3SK hynix — SK hynix Announces 3Q26 Financial Results
news.skhynix.com
- 4
- 5
- 6
- 7
- 8Axiom — Formalizing the Enormous Theorem
axiommath.ai
- 9Axiom — CFSG in Lean (project page)
cfsg.axiommath.ai
- 10
- 11
- 12
- 13
- 14GitHub — AxiomMath/CFSG
github.com
- 15Hexagon — The Unbalanced Group Theorem
hexagonmath.org
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.