# The Frontier — Edition No. 37

> Monday, November 2, 2026. 7 items from named primary sources, with independent analysis.

## 1. Marvell will buy Lumentum for $115 billion in stock, betting data centers switch to light

Source: [The Information](https://www.theinformation.com/briefings/marvell-buy-lumentum-115-billion-all-stock-deal)

Chipmaker Marvell agreed on Monday to buy laser maker Lumentum in an all-stock deal valuing Lumentum at $115 billion, one of the largest semiconductor mergers ever. Marvell designs chips for data centers, networking, and storage, and is buying a leader in lasers and other optics, the light-based parts that move data inside data centers. The deal is a bet that data center wiring will shift from copper cables to optical and photonic technology, which uses light instead of electricity to carry signals. Marvell had already moved in this direction with its purchase of startup Celestial AI earlier this year. If the deal closes, Marvell shareholders will own 67% of the combined company, Lumentum shareholders the rest, and Marvell CEO Matt Murphy will lead it. Approval is not certain, and a deal this large is expected to face close review, especially in China, where Marvell does significant business. Nvidia invested $2 billion in each company this year.

**The read:** This is a bet on the plumbing inside AI data centers, not on a new model. As AI systems grow, moving data between thousands of chips becomes the bottleneck, and copper wires run out of speed and energy headroom faster than light-based links do. Whoever owns both the chips and the light-based connections can sell the whole package, which is why a $115 billion price makes sense to Marvell even before regulators weigh in. For hardware founders, the lesson is that the next choke point is not computing power alone, but how fast and cheaply data travels between machines.


## 2. Anthropic revenue hit nearly $17 billion in Q3, but it swung back to an operating loss ahead of its IPO

Source: [The Information](https://www.theinformation.com/briefings/anthropics-revenue-surged-17-billion-q3-swung-back-operating-loss)

Anthropic reported nearly $17 billion in revenue for the third quarter ending in September, up 47% from the second quarter, in an updated filing on Monday morning ahead of a blockbuster IPO expected this month. Growth is still fast but slowing from its earlier pace. The company reported a $23.3 billion net loss for the quarter, nearly double the $12.1 billion net loss a year earlier, which it linked to $5 billion in costs to secure short-term computing capacity and $2 billion in donations matching those made by employees. It swung to a $567 million operating loss after a $974 million operating profit in the second quarter, though it reported $2.5 billion in adjusted operating income once stock-based pay and other costs are set aside. It did not disclose a revenue run rate as of the end of October, even though many investors expected one. The filing also shows $895 billion in commitments for computing and data center leases.

**The read:** Anthropic is now big enough that its numbers read like a public company, and the story is the gap between sales and spending. Revenue is still growing fast, yet the cost of renting enough computing power to serve it, plus large one-off items, pushed it back into an operating loss. The $895 billion in future lease commitments is the number to watch: it locks in years of computing bills before the revenue to pay for them is certain. For investors weighing the IPO, the question is no longer whether customers want the product, but whether the computing bill grows slower than sales.


## 3. OpenAI signed ad-tech deals with Magnite and PubMatic to fill unsold ChatGPT ad space

Source: [The Information](https://www.theinformation.com/briefings/exclusive-openai-sign-deals-magnite-pubmatic-fill-unused-chatgpt-ad-space)

OpenAI has signed deals with Magnite and PubMatic, companies whose software helps advertisers place ads on websites and other digital platforms, to fill unused ad space on ChatGPT, according to people familiar with the deals. Advertisers will use the two companies' tools to set budgets, place bids, and upload ads, while OpenAI's own system keeps control of where and how ads are shown, an OpenAI spokesperson said. The partnerships have not been announced, and the volume of ads flowing through them is still small, though the ad companies expect it to grow quickly. OpenAI has struck similar deals with Criteo, LiveRamp, and StackAdapt, and with Amazon Ads in September, and it has cut ad rates and added tools such as ads that let a user chat with a brand by clicking.

**The read:** ChatGPT ads are moving from an experiment to normal ad inventory that any advertiser can buy through tools they already use. That matters because the hardest part of an ad business is not showing ads, but getting thousands of advertisers to spend without learning a new system. By plugging into Magnite and PubMatic, OpenAI removes that friction, which should grow ad revenue faster than building its own sales channel alone. For publishers and app makers, the warning is clear: the most valuable new ad space is inside the chatbot, and the companies that control its placement rules will set the price.


## 4. DeepSeek showed a prototype coding agent that watches group chats and fixes problems on its own

Source: [The Information](https://www.theinformation.com/briefings/deepseek-unveils-autonomous-coding-agent-prototype)

DeepSeek's Harness team open-sourced a prototype it calls a 'super agent' over the weekend, as the Chinese AI lab steps up competition with Anthropic's Claude Code, OpenAI's Codex, and other coding tools. The team stressed it is a research prototype, not an official DeepSeek product, built to need as little human help as possible. When it takes on a task, it can watch feedback in public group chats, find the root cause of a problem itself, and propose a fix. People can talk to it across messaging apps, GitHub, and ByteDance's workplace app Feishu: if a user first tells it one method in a chat app and later asks on GitHub to switch methods, it can follow the change and carry out the task. It is also built to juggle several complex tasks at once, so people spend less time managing sessions and more time reviewing and approving the final work.

**The read:** The coding-agent race is moving from answering a single prompt to living where work actually happens. An agent that watches a team chat, notices a bug report, and starts fixing it without being asked is far more useful, and far harder to control, than one that waits for instructions. DeepSeek is also showing its work in public, which lets other builders copy the design fast. For teams that write software, the practical question is shifting from which coding tool to buy to how much unsupervised work you are willing to let an agent do, and what review step catches its mistakes before they reach customers.


## 5. Salesforce will buy defense startup Rune for a reported $375 million, extending its buying streak

Source: [The Information](https://www.theinformation.com/briefings/salesforce-buy-defense-tech-startup-extending-acquisition-streak)

Salesforce said on Friday it will buy Rune Technologies, a two-year-old defense startup founded by two former Anduril Industries employees, to build out Missionforce, its suite of software tools for the military. The price was reported at $375 million in cash and stock by the publication Resilience Media. Rune sells software that helps agencies such as the US Marine Corps and US Army forecast supply needs and handle logistics, had raised $30 million from investors including Andreessen Horowitz and Human Capital, and was last valued at $134 million, according to the same publication. It is the latest in a run of Salesforce deals this year: AI customer-service agent Fin for $3.6 billion in September, a planned $5.7 billion purchase of customer-engagement company Braze announced last month, customer-research firm Listen Labs, at least $1 billion for Germany's Contentful, and Israeli analytics company Loops.

**The read:** Salesforce is buying its way into defense software, a market where contracts are long, budgets are large, and trust matters more than a slick demo. Rune is small, but it already sells to the Army and Marines, which gives Salesforce a foot in the door that would take years to earn from scratch. For founders in government and defense tech, this is the exit path taking shape: build a working product for one hard military job, prove it with real users, and a big software company will pay several times your last valuation to own that relationship. For investors, the streak itself is the signal: Salesforce is choosing to buy growth in new markets rather than wait for its core business to speed up.


## 6. Binance saw $5.8 billion leave after withdrawals resumed, about 3% of its assets

Source: [The Information](https://www.theinformation.com/briefings/binance-allows-customers-resume-withdrawals-sees-5-8-billion-net-outflow)

Binance, the world's largest crypto exchange, saw $5.8 billion in net outflows after it resumed processing withdrawals on Sunday morning following the $6.3 billion hack last week. The figure, tracked by crypto analytics platform DefiLlama, is about 3% of Binance's $171 billion in assets under management before the hack. Binance said withdrawal volumes spiked at first and then declined and stabilized by Sunday midday. Bitcoin traded around $77,700 over the weekend, flat after falling from $80,000 before the hack on Thursday. The FBI has said North Korea was behind the theft, and Binance has pledged to cover all losses.

**The read:** A $5.8 billion outflow sounds alarming, but at 3% of assets it is a vote of limited panic, not a bank run. Most customers stayed put, which suggests Binance's pledge to cover every loss, plus the speed of restoring withdrawals, worked as damage control. For crypto founders, the test that matters now is slower: whether big traders and institutions quietly move funds over the next few weeks once the headlines fade. If outflows stay near this level, the hack becomes a costly security story; if they keep climbing, it becomes a trust story, and trust is much harder to buy back.


## 7. Modal's CEO on 20VC: who wins the inference market, and whether open-source inference has peaked

Source: [20VC](https://www.youtube.com/@20VC)

On 20VC, Modal co-founder and CEO Erik Bernhardsson discussed who wins the market for inference, the business of running trained AI models for customers, over the next five years, and whether open-source inference has already peaked. Modal, which provides infrastructure for training models, running inference, and powering AI agents, has passed $1 billion in annual recurring revenue after growing more than 15 times in 12 months, and raised over $1.2 billion in total, including a recent $750 million Series D led by Accel at a $15.75 billion valuation. Despite that growth, its cash flow is still negative because it keeps spending on computing capacity. The episode also covers the real limits of post-training and reinforcement learning, the methods used to improve a model after its first training run, and whether 90% software profit margins are dead in AI.

**The read:** Modal's numbers show where the money in AI is moving: from building models to running them, day after day, for real customers. Growing 15 times in a year to $1 billion in revenue, while still losing cash to buy computing power, is the same trade Anthropic is making at a larger scale. The open question Bernhardsson raises is whether anyone can keep high software-style profit margins when the product is mostly rented computing power. For founders picking where to build, infrastructure that makes running models cheaper and more reliable may be a better business than yet another model, but only if you can survive the capacity bill while you scale.


[View this edition](https://thefrontier.news/editions/2026-11-02)