# The Frontier — Edition No. 38

> Tuesday, November 3, 2026. 10 items from named primary sources, with independent analysis.

## 1. Salesforce will buy AI contact-center startup Cresta for $2.7 billion

Source: [The Information](https://www.theinformation.com/briefings/exclusive-salesforce-acquires-startup-cresta-2-7-billion)

Salesforce announced Monday it plans to acquire Cresta, a nine-year-old company that automates contact centers with AI. Salesforce did not disclose a price, but agreed to pay $2.7 billion, according to a person familiar with the deal. Cresta began as a Stanford AI Lab project, was founded in 2017 by Sebastian Thrun, Tim Shi and Zayd Enam, was valued at $1.6 billion after its 2022 Series C, and had raised more than $280 million from Andreessen Horowitz, Sequoia Capital and Tiger Global. It is Salesforce's second big customer-service purchase in two months, after the $3.6 billion Fin deal closed in September, and it sharpens competition with startups such as Sierra, led by former Salesforce co-CEO Bret Taylor.

**The read:** Salesforce is buying the company that answers your support calls, not just the software that tracks them. Customer service is one of the first jobs where AI agents can do the whole task, so owning the best automation there is worth more than adding another dashboard. For founders, the signal is that the acquirers are paying up for products with real enterprise customers and proven savings, not for demos. For investors, the price, about 70% above Cresta's last valuation, shows that proven AI revenue still commands a premium even in a cautious market.


## 2. Oracle raises $20 billion in stock and bonds as AI spending strains its balance sheet

Source: [The Information](https://www.theinformation.com/briefings/oracle-launches-20-billion-stock-bond-sales)

Oracle filed Monday to raise $20 billion, split evenly between investment-grade bonds and common stock, to fund its heavy spending on AI infrastructure. Shares fell more than 9% on Monday, deepening its loss for the year to more than 39%. Oracle had previewed the plan last week, saying it would limit further capital raising to $20 billion total through fiscal 2030 and that free cash flow would turn positive in fiscal 2028, which begins in June 2027. The move follows large stock sales by Meta and a $25 billion bond sale by Microsoft last week.

**The read:** The AI buildout is now being paid for in public, with real dilution and real debt, and shareholders are keeping score. Oracle is telling the market it can cap its fundraising and reach positive cash flow by 2028, but the 9% share drop says investors want proof, not a plan. Watch the sequence: Microsoft borrows, Meta sells stock, Oracle does both. The builders who can show when spending turns into cash will keep raising cheaply; the ones who cannot will pay more for every dollar.


## 3. Anthropic pitches IPO investors on Claude designing its own AI chips

Source: [The Information](https://www.theinformation.com/articles/anthropic-touts-claudes-chip-design-skills-ipo-investors)

Ahead of an initial public offering later in November that could value it around $1.5 trillion, Anthropic has told investors its new 15-person chip design team is already using Claude models for engineering and design work on its own custom chips, cutting design time compared with traditional workflows. Anthropic announced in August it would design its own chips to train and run Claude, after talks with Samsung as a possible manufacturing partner. OpenAI is on the same path with its Jalapeno chip, developed in nine months with Broadcom for significant use next year. Claude's top model scored best on a ChipAgents and Nvidia chip-design test last month, but only reached 16.6%. Anthropic has not said it will sell Claude for chip design, and its biggest backers, Amazon, Google, Nvidia and AMD, have committed about $90 billion to it combined, which could create tension if it competes with them.

**The read:** The model labs are no longer just customers of the chip industry; they want to become chip designers, and they are using their own AI to do it. A 16.6% test score shows the technology is early, but the direction is what matters: if AI can design chips, the cost and speed of building custom hardware fall for everyone, and the specialist design software firms, Synopsys and Cadence, face a new kind of rival. The awkward part is the funding: Anthropic's biggest backers are also the chip companies it would compete with, so its pitch to investors is also a negotiation with its own shareholders.


## 4. How smugglers get expensive memory chips into China, and why it matters for Huawei's next AI chip

Source: [The Information](https://www.theinformation.com/articles/smugglers-get-expensive-memory-chips-china-south-korea)

United States rules since late 2024 ban shipping high-bandwidth memory, the fast memory stacked next to AI processors, to China on its own. But the rules allow that memory when it is packaged with a cheap server processor below the export threshold, because the package is judged by the processor. Smugglers order top Samsung and SK Hynix memory, package it with deliberately weak processors through firms in Taiwan, South Korea and Southeast Asia, declare it under the processor's customs code, ship it to Hong Kong, then carry it into mainland China in luggage, car trunks or commuters' bags. There, technicians heat the packages until the solder softens and pry the memory stacks off, discarding the processors. Huawei has told some clients that a large share of its next flagship AI chip, the Ascend 960DT due in the first quarter of next year, will run largely on Samsung and SK Hynix memory because domestic suppliers cannot make enough. A single high-bandwidth memory unit now sells for about $4,000 in China, up from several hundred dollars a year ago, and a chip package needs six to eight of them, adding $24,000 to $32,000 for memory alone.

**The read:** Export controls are being beaten not by better smuggling boats, but by a paperwork loophole: judge the package by its weakest part, and the valuable part rides along for free. That loophole is now priced into the market, with memory selling for roughly ten times last year's price inside China. For anyone building or funding AI hardware, the lesson is that supply chains, not just chip designs, decide who can train frontier models. And for policymakers, every gap like this one teaches the next workaround, so controls that are not airtight mostly raise the price of getting caught, not the price of getting through.


## 5. Nvidia tells cloud customers to take on less China business

Source: [The Information](https://www.theinformation.com/briefings/nvidia-tells-cloud-clients-take-less-china-business)

Nvidia has recently told some cloud computing customers to reduce their exposure to Chinese companies, according to three people familiar with the matter, in a move that could narrow a key route for China's AI developers to reach advanced computing power by renting it overseas instead of importing chips. The guidance aims in part to push cloud providers, including Nvidia-backed newer cloud firms, toward a more diverse customer mix. It is not mandatory: customers are asked to submit the identities of the end users of the chips, and it has not so far been enforced as a hard rule against taking Chinese customers.

**The read:** The second front of the chip war is not hardware, it is rented computing power. If Chinese developers cannot buy chips, they rent them abroad, and Nvidia is now asking its cloud customers to police that route themselves. Because the guidance is not yet a hard rule, the real question is what happens the first time a big customer refuses: does Nvidia enforce it and lose revenue, or look the other way and lose credibility with regulators? Cloud providers caught in the middle should expect customer screening to become a cost of doing business, not an optional favor.


## 6. OpenAI's LifeVal grades AI on everyday life, not exams

Source: [OpenAI](https://openai.com/index/lifeval/)

OpenAI published LifeVal on November 2, a test that measures how well AI handles practical everyday tasks instead of academic questions. The tasks were chosen from United States and international time-use surveys plus a new survey of 3,995 people, covering caregiving, health, education, household paperwork, personal finances, shopping and government services. OpenAI says its latest models score very highly on the showcased tasks, including 93.1% for comparing kindergartens after a move, 98.0% for organizing flood-preparation volunteers, and 100% for finding a home that fits a budget, commute and separate living unit. The company says a year of work on browser use, computer use and software connectors drove the improvement, moving ChatGPT from a chat tool toward an assistant that takes actions.

**The read:** The benchmark race is moving from can the model pass a hard exam to can it actually get your life admin done. That shift favors whoever owns the messy middle: permissions, logins, forms and follow-through, not just raw intelligence. If everyday tasks become the scoreboard, products that connect to real services will look better than models that only answer well, and startups building on top of these assistants should expect the assistant itself to swallow more of their workflow. The honest caveat is that these are OpenAI's chosen examples, so independent tests on the same tasks will matter more than the headline scores.


## 7. Enterprise customers are refusing bigger AI contracts at renewal time

Source: [The Information](https://www.theinformation.com/newsletters/ai-agenda/contract-renewal-season-customers-biting)

As annual contracts come up for renewal, enterprise customers are resisting larger commitments to Anthropic, OpenAI and other model providers, according to The Information's AI Agenda newsletter. Two reasons stand out: new models arrive so fast that customers want the freedom to switch week to week, and providers keep releasing cheaper models that are good enough, such as OpenAI's Terra and Luna and Anthropic's Haiku, which can cut bills to a fraction of earlier levels. One example cited is Replit, which switched its product from mostly Anthropic to OpenAI models in a matter of hours. Customers are turning instead to model routers such as OpenRouter and multi-model platforms such as Snowflake. That sits uneasily with Anthropic's prospectus goal of more multi-year, non-cancelable contracts with minimum spending, though Anthropic says its third-quarter net retention rate was 900%.

**The read:** The labs want customers locked in for years; customers want to stay free to switch every week, because the cheaper model next month may do the same job for a tenth of the price. That tug-of-war is the most important pricing dynamic in AI right now. If routers and multi-model platforms win, models become a commodity you buy by the cheapest unit, and the winners are the companies that route, manage and guarantee quality across models, not the ones that train a single model. Anyone selling AI to businesses should stop pitching lock-in and start proving they are worth keeping when the customer can leave in an afternoon.


## 8. SpaceX is in talks with Boost Mobile to become a real mobile carrier

Source: [The Information](https://www.theinformation.com/briefings/spacex-talks-boost-mobile-expand-wireless-coverage)

SpaceX is in talks with discount carrier Boost Mobile to expand its wireless coverage beyond its satellite network, according to The Washington Post as reported by The Information. A deal could mean acquiring the mobile network, which is owned by EchoStar, or gaining access to the network it leases from AT&T and T-Mobile. Elon Musk wants ground-based cellular coverage to supplement Starlink satellites, which struggle indoors, in dense cities, or without a clear view of the sky. SpaceX previously tried deals with Verizon, AT&T and T-Mobile, but they turned it down, and carriers view Musk's wireless ambitions as a threat.

**The read:** Starlink proved satellites can cover the countryside, but phones live indoors and in cities, where satellites are weak. Buying or renting a ground network would turn SpaceX from a satellite company into a full mobile carrier, competing directly with the firms it once courted as partners. For telecom investors, the question is whether regulators and carriers let a company that already dominates satellite internet also own a slice of the ground network. For everyone else, it is a reminder that distribution, owning the pipe to the customer, is still the prize, even in the space age.


## 9. Apple's sales cool, with services growing at their slowest pace in over three years

Source: [The Information](https://www.theinformation.com/briefings/apple-sales-cool)

Apple's September-quarter revenue was $114.5 billion, up 12% from a year earlier, its slowest sales growth of the year, in its first earnings under new CEO John Ternus. The iPhone stayed strong, up 17% to $57.6 billion, but iPad sales fell 10% and wearables fell 8%. Services, which includes iCloud, the App Store and Apple TV, grew slightly less than 10% to $32 billion, its slowest quarterly growth in more than three years, and Apple projects services growth of 8% to 9% in the December quarter. Apple expects total sales to grow 7% to 10% in the current quarter, and shares fell nearly 3% in after-hours trading.

**The read:** Apple's growth engine is splitting in two: hardware still sells, but the high-margin services business that investors counted on for steady growth is slowing. Services slowing to single digits matters because it was supposed to be the reliable part of the story while iPhone sales rose and fell. For app developers and subscription businesses, slower App Store and services growth means a tougher fight for the same customer dollar, and for investors it raises a plain question: if neither phones nor services are accelerating, what is Apple's next growth story?


## 10. a16z Show: the robots already working in American factories

Source: [a16z](https://a16z.simplecast.com/)

On the November 2 episode of the a16z Show, Tutor Intelligence CEO Josh Gruenstein joined general partners Martin Casado and Raghu Raghuram to discuss what it takes to build robots that actually work in factories and warehouses. Gruenstein's argument is that robotics is a systems problem spanning AI, hardware, manufacturing and the work itself, not a single breakthrough. Tutor's approach is to deploy robots into real businesses, use human tutors when the robot gets stuck, and turn that human help into training data that makes the robot more capable over time.

**The read:** The factory robot race will be won by whoever handles failure gracefully, not whoever demos best. A robot that stops and waits for a human is useless on a production line; a robot that learns from each human rescue gets better every shift. That turns deployment into the product: the company with the most real factory hours, and the best system for turning human fixes into robot skill, builds a lead that a lab demo cannot copy. For founders in robotics, the takeaway is blunt: sell working hours in a real building before you sell a vision of full autonomy.


[View this edition](https://thefrontier.news/editions/2026-11-03)