# The Frontier — Edition No. 32

> Wednesday, October 28, 2026. 10 items from named primary sources, with independent analysis.

## 1. Amazon lists 250 million rival products to keep shoppers away from AI agents

Source: [The Information](https://www.theinformation.com/articles/amazon-lists-250-million-products-rival-sites-effort-battle-agents)

Amazon's Shop Direct program now lists 250 million products sold elsewhere, about 2.5 times as many as in March, as Amazon tries to keep shoppers from turning to outside AI shopping agents such as Meta's Muse. Amazon says tens of millions of customers have found hundreds of millions of products from other stores through Shop Direct, which appears in hundreds of millions of searches a month, and it has added brands such as Alo Yoga and Warby Parker. The push builds on outside listings launched in February 2025 and a Buy for Me feature added two months later, and it matters for Amazon's $588 billion commerce and advertising business. Merchants are included unless they opt out, yet sellers representing hundreds of merchants said they have little insight into when or why products appear and have seen few sales from the listings. Amazon has blocked outside agents from its site for more than a year.

**The read:** Amazon would rather show you a rival's product itself than let an AI agent take you to that rival. If agents become the new storefront, the company that owns the shopping trip owns the sale, so Amazon is building a bigger shelf to stay in the middle. For sellers, the deal is mixed: free exposure on the biggest store, but little control and few sales so far. For agent builders, the warning is plain: the biggest retailer will fight to keep the customer, not hand them over.


## 2. Chipmaker Onsemi is in talks to buy power startup Ferric for about $1 billion

Source: [The Information](https://www.theinformation.com/articles/chipmaker-onsemi-talks-buy-startup-1-billion)

Onsemi has discussed buying Ferric, a 16-year-old New York chip startup, for close to $1 billion, with a deal possibly coming together in the coming days, though terms could change or the talks could fall apart, and the price could include payments tied to Ferric hitting performance targets. Ferric makes integrated voltage regulators that send power directly to small chips underneath or inside AI processors, cutting energy lost on the way. The talks follow Analog Devices' May agreement to pay $1.5 billion for Empower Semiconductor. Scottsdale-based Onsemi has about a $30 billion market value and $6 billion in revenue last year, and this month it agreed to buy Synaptics for $5.7 billion in cash.

**The read:** The AI bottleneck has moved inside the chip itself, to the tiny job of delivering power without wasting it. As data centers draw more electricity, the company that feeds power cleanly to the processor gains an edge, and big chipmakers are now buying that skill instead of building it. Expect more deals in this quiet corner of the chip world, because power delivery, not raw computing speed, may decide which AI hardware wins.


## 3. OpenAI says its models produced three major science results, and hands the checking to outside experts

Source: [OpenAI](https://openai.com/index/building-scientific-understanding-with-academic-communities/)

On October 26, OpenAI announced three unreviewed claims from its internal models: a proof of a 1991 wave-math conjecture in every dimension, a proof that the Exponential Time Hypothesis is false, and a proof that certain quantum field theories, including the theory of the strong force, exist as rigorous mathematical objects. OpenAI calls the manuscripts raw research data, not settled science. Outside expert groups will examine them, led for the computer science claim by the hypothesis's original proposers, Russell Impagliazzo and Ramamohan Paturi, with Ryan Williams. The math and computer science proofs and their computer-checked versions are public; the physics materials are held back while its expert group, led by Nobel laureate David Gross, writes its own account. These claims are unconfirmed until independent experts verify them.

**The read:** If even one of these claims holds up, it changes what AI is for. Science would move from humans using AI as a helper to AI producing results that humans then check, which makes checking, not generating, the scarce skill. For founders, the near-term lesson is smaller and more useful: any product that can prove its own work, step by step, will be trusted faster than one that only sounds right. Watch the expert groups, not the headlines, because their verdict decides whether this is a breakthrough or a warning about AI-generated proofs.


## 4. Anthropic buys a one-person drug startup and reports early results across six drug programs

Source: [Anthropic](https://www.anthropic.com/research/early-progress-drug-development)

Anthropic said on October 27 that it has acquired Mithril Medicine, a one-person company developing treatments for two rare forms of muscular dystrophy, and that its founder, Cody Tranbarger, used Claude to organize contracts, logistics, and experiments with 11 outside labs, taking a candidate from idea to promising results in mice in under 18 months. A drug based on that work could enter human trials as soon as late 2028, though Anthropic stressed the work is mostly preclinical. Anthropic also reported deep research partnerships with six organizations, including Bristol Myers Squibb, Generate Biomedicines, BigHat Biosciences, Divergence Labs, Enveda, and The Column Group, with early results including a cancer-therapy design about 70% shorter, 35,000 antibody designs in about a month, and paperwork review time cut 89%.

**The read:** A leading AI lab just became a drug developer, not just a tool seller to drug developers. That matters because drug development is slow and expensive mainly due to coordination, paperwork, and failed designs, which are exactly the jobs an AI agent can take on. If one person plus AI can run a program that once needed a whole company, small biotech teams get far more powerful, and big pharma companies that move slowly will feel the pressure. The catch is that mouse results are not human results, so the real test arrives in 2028, not today.


## 5. OpenAI and Oracle bring business software into ChatGPT

Source: [OpenAI](https://openai.com/index/oracle-apps-in-chatgpt/)

OpenAI and Oracle announced on October 27 that Oracle applications and data will come into ChatGPT, and that new Oracle products will be powered by OpenAI models. Oracle is developing Oracle AI Workspace, which would let employees describe a task and complete approved business workflows across Oracle applications from inside ChatGPT, within each company's permissions and approvals. Oracle reported 130,000 active ChatGPT users and more than 95,000 active Codex users inside Oracle. The deal also adds Oracle AI Security Mesh, Oracle App Composer, and an Oracle AI SMB Bundle. Customers in the United States can buy the OpenAI API, Codex, and ChatGPT Work through Oracle Marketplace using eligible Oracle Universal Credits.

**The read:** ChatGPT is moving from answering questions to doing the job inside the software companies already run on. That is a big distribution win for OpenAI, because Oracle's customers do not need a new buying process, they just add AI to a bill they already pay. For startups selling business software, the bar just rose: your product now competes with a chatbot that can reach into the customer's finance, hiring, and planning systems. The whole deal rests on permissions, so the winners will be the teams that prove an agent only touches what each employee is allowed to touch.


## 6. Anthropic cuts Claude prices 40% for government buyers at every level

Source: [Anthropic](https://www.anthropic.com/news/expanding-partnership-gsa)

Anthropic expanded its agreement with the U.S. General Services Administration on October 27, giving government buyers a 40% discount on Claude Enterprise, Claude for Government, and the Claude API. For the first time, the discount covers every product surface, applying to direct purchases now and to purchases through cloud providers in the coming months, and it extends beyond federal agencies to state, local, tribal, and territorial governments. Agencies pay only for what they use, with no minimum spending commitment, and Anthropic estimates the discounts will save the government $1 billion. Anthropic says about 500 U.S. federal, state, and local government organizations already use Claude.

**The read:** The government is becoming a real AI customer, not just a regulator watching from the side. A 40% discount with no minimum spend removes the two excuses that slow public-sector buying, price and commitment, and extending the deal to states and cities opens thousands of new buyers at once. For AI companies, government deals are now a distribution channel worth fighting for, and the company that wins the most agencies gets a steady, sticky revenue base that does not depend on consumer fads. Expect rivals to match the discount, because losing the government market early is hard to undo later.


## 7. Goldman banker: AI is not a bubble, but bubbles are forming inside it

Source: [The Information](https://www.theinformation.com/briefings/goldman-sachs-investment-banking-co-head-sees-repositioning-amid-ai-boom)

Kim Posnett, Goldman Sachs' global co-head of investment banking, said at The Information's WTF Summit on October 27 that areas of weakness are emerging in the broader AI market as the cost of capital is set to keep rising, even though she remains hugely bullish on AI overall. Investors are questioning what higher interest rates could mean for AI capital spending and financing, and AI projects will not deliver expected returns if rates stay high for a long period. She called the shift a repositioning and differentiation across the AI trade, not a change in feeling about AI, adding that while AI itself is not a bubble, bubbles can exist within it. Investors are starting to favor vertically integrated companies that produce frontier models, software, and AI chips, while Goldman is playing leading roles in SpaceX's GPU financing, Meta's stock offering, and Anthropic's planned IPO.

**The read:** The banker at the center of AI's biggest deals just gave founders a sorting test. Money is not leaving AI, but it is getting pickier, favoring companies that control more of the stack, from models to chips to software, over companies that rent one layer of it. If your business depends on cheap capital or on a giant customer who is also your investor, expect harder questions. The firms that can show real demand, not just impressive funding rounds, will be the ones that keep raising.


## 8. Bumble goes private at $1.3 billion, down from a $14 billion IPO value

Source: [The Information](https://www.theinformation.com/articles/bumble-founder-whitney-wolfe-herd-explains-decision-take-company-private)

Bumble cofounder and CEO Whitney Wolfe Herd said at The Information's WTF Summit, in her first interview since the sale was agreed, that taking the 12-year-old dating app private mattered for the longevity of its mission after its stock dropped to near $2. Ithaca Holdings agreed earlier in October to take Bumble private in a $1.3 billion deal, down sharply from its 2021 IPO at roughly a $14 billion valuation; its market value has since sunk to under $1 billion amid falling revenue and paying users. Wolfe Herd will step down as CEO, handing the role to Ithaca Holdings founders Ravi Gupta and John Hegeman as co-CEOs, while remaining involved as an active shareholder. She blamed fatigue among users in their 20s and 30s who are overexposed to technology, but said AI, including a new assistant called Bee, could modernize the product.

**The read:** A company can lose 90% of its public value and still be worth buying, if the buyer believes the product can be rebuilt in private. Bumble's story is a warning for consumer apps: when users tire of the feed, no growth trick saves you, and the fix has to be a better product, not a better pitch. For founders, the lesson is that going private is not failure, it is a reset button, but only if the new owners have a real plan for AI to make the product feel new again.


## 9. Malaysia seizes $12 million in AI chips in a smuggling crackdown

Source: [The Information](https://www.theinformation.com/briefings/malaysian-authorities-seize-12-million-worth-ai-chips-smuggling-crackdown)

Malaysian customs seized about 49.5 million ringgit ($12.11 million) worth of advanced AI chips this month while investigating a syndicate suspected of routing chips through the country's main airport. Authorities found 30 server units containing advanced AI chips at Kuala Lumpur International Airport's free trade zone; the shipment, from one Asian country and bound for re-export to another, was falsely declared as electronic components to bypass restrictions. Malaysia tightened export controls last year on advanced U.S.-origin chips following pressure from the Trump administration to curb China's access to AI-critical semiconductors, and at least five cases linked to the syndicate involve seizures worth more than 100 million ringgit.

**The read:** Chip controls are now enforced at airports, not just in trade documents. A $12 million seizure is small next to the AI buildout, but it shows that smuggling routes through third countries are being watched and closed, which raises the cost and risk of getting restricted chips. For hardware builders, the message is to treat supply-chain paperwork as seriously as engineering: one mislabeled shipment can strand a whole order and draw investigators.


## 10. OpenAI opens up data on how 165 kinds of work flow through ChatGPT

Source: [OpenAI](https://openai.com/index/chatgpt-expanding-world-of-work/)

OpenAI released data on October 26 mapping ChatGPT work use to 165 work activities across nearly 100 countries, covering the first quarter of 2025 through the third quarter of 2026. The data covers work-related messages on individual ChatGPT plans, not enterprise messages or Codex use. Across 74 countries observed in every quarter, the ten most common activities fell from an average of 56.0% of each country's classified work messages in early 2025 to 51.1% in the third quarter of 2026, which OpenAI reads as countries specializing in different kinds of work over time.

**The read:** This is the first public map of what people actually ask AI to do at work, country by country, and the trend points away from a few generic uses toward many specialized ones. That tells builders to stop chasing one killer work feature and instead serve a specific job well, because demand is spreading out, not concentrating. For investors and hiring managers, the data is a hint about which tasks get automated first in each market, and which local work habits will shape the next wave of AI products.


[View this edition](https://thefrontier.news/editions/2026-10-28)