Bill, what's a company that makes lasers worth these days? I'm guessing the answer ends in billion. One hundred fifteen of them, if you use Marvell's math. Good morning. It's Tuesday, November 3rd, and this is AI Daily Briefing. In today's rundown: Marvell agrees to buy optical-parts maker Lumentum in an all-stock deal, Anthropic's amended IPO filing warns it may be close to its own AI-acceleration threshold, Brookfield plans a data center campus at Britain's Drax Power Station, and Salesforce goes after the phone line with a deal for voice-AI startup Cresta.
Marvell and Lumentum's joint announcement on Monday, with Anissa Gardizy and Robbie Whelan reporting for The Wall Street Journal:
The transaction has an implied enterprise value of $115 billion for Lumentum, representing approximately $1,132 per share of Lumentum common stock and a premium of approximately 23.9% based on each company’s unaffected closing share price as of October 28, 2026, the last trading day before market speculation regarding a potential transaction began.
The terms: Lumentum holders get four and a half Marvell shares for every share they own, and current Marvell holders end up with about sixty-seven percent of the combined company. Lumentum brings lasers, optical transceivers and optical circuit switches. Marvell brings electrical and optical interconnect, network switching and custom silicon. The headline number depends on who you read. The Journal calls it a roughly eighty-billion-dollar deal. The companies' one hundred fifteen billion is an implied enterprise value, priced off closing prices from Wednesday, October 28th, before word of the talks got out. Per a roundup in The CODEW, Marvell fell as much as fourteen percent after Bloomberg reported those talks. In an all-stock deal the market sets the price, and the market wasn't cheering. The pitch is that copper is running out of room. Both companies argue that as AI clusters scale, connectivity has to shift to optics co-designed with the chips. They put the combined addressable market at five hundred fifty billion dollars and promise more than a billion dollars a year in run-rate cost synergies. Those are company figures. Nobody independent has checked them. Five hundred fifty billion is a slide in the investor deck, not revenue. What I'd want to know as a buyer: does one merged supplier mean pre-validated optics that ship faster, or just one fewer vendor to play off against the other? This is the infrastructure layer I keep saying gets less attention than model launches. The CODEW notes Nvidia had already put two billion dollars into Lumentum. Lumentum's SEC filing says closing should take roughly twelve to eighteen months, subject to shareholder and regulatory approvals, and analysts quoted by TechNews name review in both the US and China as a risk. Eighteen months is a long time for customers to wonder whose roadmap they're buying into.
Integuide AI News, reading Anthropic's amended filing with the SEC:
Anthropic's first amendment to its IPO prospectus, filed 2 November, adds that "our internal estimates and our expectations regarding future models, rather than our current models" indicate it "may imminently meet" its Responsible Scaling Policy's threshold of "dramatically accelerated progress attributable to AI".
Per Integuide, that threshold means progress running at roughly twice its expected pace, plausibly because AI is automating research. Think two years of progress at the 2018-to-2024 rate squeezed into one. Anthropic says no current model has been designated as crossing it. It will publish further reports under the policy, and it says its response could affect whether, how and when it develops or deploys models. Two admissions jumped out at me. Anthropic concedes that in the past it has only spotted accelerations after they happened. And it says it doesn't expect to be able to safely let Claude fully autonomously build successive frontier models. That's a lab writing its own internal forecasts into a securities filing. Internal is the key word. No outside evaluator has checked this, and risk-factor language tends to be written broadly. But it's in a filing investors will rely on, not a blog post, and it gives them a stated reason the release schedule could slow. Then there's the money. Stillflood's AI Frontier Daily, citing the amendment and Bloomberg, says Anthropic's purchase commitments and lease agreements total about eight hundred ninety-five point four billion dollars, substantially all for compute and infrastructure, with about four hundred forty-seven point five billion due before the end of 2030. The named minimums include about one hundred twenty-one billion to Google, at least one hundred ten billion to Amazon and at least forty-four point nine billion to Microsoft. When the original S-1 came out on October 26th, reports put the figure at six hundred fifty-seven billion plus. These digests don't say whether the increase reflects new contracts or a broader count, so treat that comparison with caution. Third-quarter revenue came in just under seventeen billion dollars. Anthropic also registered in Japan to sell up to three point five billion dollars of stock to Japanese investors, with Mizuho Securities leading. And a newsletter cited by The CODEW, quoting IFR, says the roadshow could start the week of November 9th, with a debut around November 20th. Anthropic hasn't confirmed any of that. So the pitch is: revenue is exploding, the compute bill is nearly nine hundred billion, and our own models might speed up research enough that we have to hit the brakes. A prospectus with a cliffhanger.
Kai Nicol-Schwarz, reporting for CNBC:
Brookfield estimates that total investment, including semiconductors, could exceed £20 billion ($26.8 billion), with an initial £2.5 billion investment as part of phase 1.
Drax announced the partnership this morning with Brookfield and Radiant, Brookfield's UK-headquartered AI infrastructure platform, on the Drax Power Station estate in North Yorkshire. The money comes through Brookfield's joint venture with Qatar-based Qai. Phase one targets up to a hundred megawatts, running from the first quarter of 2028. Phase two aims to ramp up to five hundred megawatts between 2028 and 2031, and CNBC says the site is ultimately looking at more than a gigawatt. Drax shares were up about eight percent in early London trading. The site is what makes this interesting. A thousand acres, four gigawatts of grid access, and two point six gigawatts of active dispatchable generation with cooling already on site. With Europe fighting high energy and land costs, per CNBC, that's a head start. Read the conditions, though. Drax says phase one depends on planning, on regulatory and network approvals from the government and the system operator, and on securing a suitable tenant. No tenant has been named. A community consultation starts shortly, and the formal planning process comes after. Compare that with Hut 8's Texas campus from yesterday's show. There, an investment-grade tenant was signed first and the bond money followed. Here the twenty-billion-pound figure comes first and the customer is still on the to-do list. And that twenty billion is Brookfield's estimate for the full build, chips included, if every phase happens. The politics are carrying a lot of weight. The release comes with a stack of ministerial quotes, the companies say the plan fits the government's new Data Centre Covenant, and science and investment minister Chris McDonald called it a vote of confidence from Qatar and Brookfield. It's the sovereign-compute pitch again, paid for with Gulf and North American capital.
Dealroom News, on Salesforce's latest acquisition:
Salesforce has signed a definitive agreement to acquire Cresta, a nine-year-old maker of AI-powered contact center software. The company didn't disclose terms, but the price is $2.7 billion, according to a person familiar with the deal.
Cresta builds voice agents that aim to resolve complex customer calls end to end, coaches human reps in real time and analyzes every interaction. Customers include Marriott, Comcast and Alaska Airlines. The two-point-seven-billion price comes from a single unnamed source, and Salesforce hasn't confirmed it. This is the latest in a string of Salesforce deals for the customer-service stack. Per Dealroom, it completed its roughly three-point-six-billion-dollar Fin acquisition and agreed to buy Braze for about five point six six billion. Cresta adds the phone call. Kishan Chetan, who runs Agentforce and Service Cloud, says voice is where service is hardest to get right. He's right, and that's my worry. A voice agent fails live, with a customer on the line and no undo button. The number I'd ask Salesforce for is how often the agent hands a call to a human halfway through, not the demo where it rebooks a flight perfectly. On Saturday we covered Deepgram buying Daily, the company behind the Pipecat voice framework. Voice AI is consolidating at both ends: infrastructure on one side, the contact-center application on the other. Cresta says it will keep supporting the contact center and telephony platforms its customers use today, and the deal is expected to close in the first quarter of fiscal 2028, pending regulatory approval.
If this briefing saves you a morning of scrolling, a rating or a quick review in your podcast app helps other builders and investors find it. Tell people which segment earned its spot today.
Tomorrow's the day Moonshot promised the Kimi K3.1 Pro weights. I'll believe it when the download finishes. And it's Election Day in the US, so tonight's results are worth watching too. AI Daily Briefing is produced by Lantern Podcasts. We'll see you tomorrow, Wednesday, November 4th.