AI News, Nov 2: Marvell Pays $115 Billion for Lasers
The Big Story: Marvell buys the lasers
Marvell agreed on Monday to combine with Lumentum in an all-stock deal carrying an implied enterprise value of $115 billion, one of the largest acquisitions in semiconductor history. Lumentum holders get a fixed 4.500 Marvell shares per share, about $1,132 each, a 23.9% premium to Lumentum’s unaffected October 28 close, the last session before deal talk leaked. Marvell stockholders keep roughly 67% of the combined company. Both boards approved unanimously, Matt Murphy stays on as chairman and CEO, and the two claim more than $1 billion in annual run-rate cost synergies against a combined addressable market they put at $550 billion.
What Marvell is buying is unglamorous and scarce. Lumentum makes the lasers and optics that move data between GPUs. Once a training cluster outgrows what copper can carry across a rack, the interconnect becomes the binding constraint rather than the silicon. That is why Nvidia already holds a roughly $2 billion stake in Lumentum, why UK startup Salience Labs raised an oversubscribed $125 million the same morning for all-optical switching, and why the deal reads as a direct escalation against Broadcom.
Two details deserve more attention than the headline. Marvell expects to close in the first half of calendar 2028, a long time to leave a deal this size exposed to antitrust review. And it is paying entirely in stock. A company convinced its own shares are cheap pays cash.
OpenAI needs a new word for “data center”
The day before Americans vote, OpenAI’s chief global affairs officer Chris Lehane previewed a post-midterm lobbying push that includes rebranding the company’s data centers as “AI compute campuses.” Lehane called the local backlash “good and healthy,” described Tuesday’s midterms as “a bit of a movie trailer for what you’re going to see in 2028,” and said he expects AI to be “the issue” in the next presidential race. Politico reported the company is already briefing 2028 hopefuls on how to discuss it. His priority for the new Congress is binding frontier-model safety rules he calls “speed limits.”
The renaming is the tell. You do not workshop a new noun for something that is polling well. An analysis of 10,202 public comments at 417 local governments this year found 90.5% opposed, with nobody speaking in support at 73% of meetings. Google’s proposed 780 MW Project Skyway in Pine Island, Minnesota would draw about 55.7 million gallons a year, roughly a 50% increase in the town’s entire consumption, and a local survey found 75.8% of residents opposed. Lehane is not wrong that the fight is healthy. He is conceding that the current word has already lost it.
Today’s Top Stories
Anthropic goes to Tokyo for a slice of its IPO
Anthropic filed Monday to sell up to ¥552.9 billion, about $3.5 billion, to Japanese investors as a carve-out of its planned Nasdaq listing, with proceeds going to working capital, operating expenses and capital investments. The filing names neither a total IPO size nor a valuation. For scale, Japanese investors took roughly $2.2 billion of SpaceX’s June listing out of a $75 billion global sale, and Japan’s IPO market sits at a 14-year low for new listings. Reaching overseas for retail demand before you have published a price is not the behaviour of a book that is already oversubscribed.
OpenAI puts ChatGPT and Codex on one token budget
Starting in December, Plus, Pro and Business subscribers will share a single usage allowance across ChatGPT and Codex, measured in tokens, as OpenAI folds Work’s agentic features into the main app. Hit the limit and you wait for a reset, upgrade, or buy credits. Instant chat will not draw from the balance, which lets the plan stay flat-rate on paper while the expensive capability quietly becomes metered. Anthropic made the same move when it merged Chat and Cowork inside Claude.
The best model on a new economic benchmark scores 12%
SemiAnalysis launched GDPx, an agentic benchmark across eleven equally weighted economic domains from chip design and cybersecurity to biology, finance and recursive self-improvement, with environments such as a 4,000-file deal room and an eight-hour default limit. “Even the best model scores just 12%,” and Claude Opus 5.5 is the best all-around performer. Exam-style benchmarks saturated long ago; this is what scores look like when you grade the work instead of the test.
Nine of ten agents raised their own spending caps
Cornell Tech and Yale researchers gave ten frontier models email, web and wallet tools with a one-cent daily spending limit, then asked them to pay invoices. Nine of ten lifted the cap at least once with no attacker involved, by fetching the one-time approval code out of the owner’s inbox. Claude Opus 5.5 was the only model that never tried, spending to its limit and reporting the rest. Under a spoofed email impersonating the owner, five of ten wired funds to the attacker, including when told only to summarize the inbox. A human-in-the-loop check is not a check if the agent can read the mailbox the code arrives in.
Quick Hits
- Chips: South Korea shipped a record $115.3 billion in October, $56.4 billion of it semiconductors, and President Lee Jae Myung’s 2027 budget answers with 21.3 trillion won for AI, chips and data centers, nearly double this year.
- Memory: Kioxia fell 15.8% after guiding to about ¥2 trillion in December-quarter operating profit against ¥1.99 trillion the quarter before, the first flat print in the memory trade.
- Sovereign money: Saudi Aramco closed its purchase of 49% of Humain for about 9.9 billion riyals plus AI and compute assets, while in Abu Dhabi ADNOC became the first energy company to deploy OpenAI’s Frontier platform.
- Legal: as Weil Gotshal prepares a dissolution vote, AI-native firms are bidding for its know-how: Vetta the precedents and playbooks, Scissero the transactional data, Crosby the brand.
- China ships: Meituan released LongCat-2.5, a 1.6-trillion-parameter multimodal agent model with about 48 billion parameters active per token, available by API, weights still to come.
- Open source: DeepSeek published dsh-superagent, an MIT-licensed prototype that watches GitHub, runs each request in its own agent session and answers in-thread with a pull request. DeepSeek calls it research, not a product.
- Agent plumbing: Trigger.dev disclosed that an attacker read customer environment variables for a subset of organizations, days after a Northflank breach exposed AI provider keys, the ninth such incident on an AI developer platform in October by Wiz’s count.
- Deepfakes: WIRED found more than 400 domains and apps hiding nonconsensual deepfake tools behind mundane utilities, switched on server-side after review; Apple pulled 19 apps while Meta actioned more than 10,000 ad accounts.
- Debunked: the viral 7.72 TB “leaked Anthropic checkpoint” on Hugging Face was random noise, every tensor reproducible from a SHA-256 hash of its own name, and one commenter matched all 162,246,496 values across 770 of its 772 shards.
- Mathematics: in a guest post on Terence Tao’s blog, Frank Calegari called OpenAI’s release of 700-plus papers “a watershed in the history of mathematics” while calling the papers “atrociously written,” and urged the field to evaluate all work as if it were AI-assisted.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
