Anthropic loses $23.3B as Marvell buys Lumentum for $115B

Monday was the day the AI boom put its accounting on the table. Anthropic amended its IPO filing and showed revenue up more than 1,000% in a year next to a $23.3 billion quarterly loss and $895.4 billion in spending already committed. Marvell paid $115 billion for a laser company, which tells you where the bottleneck moved. Oracle asked investors for $20 billion and used up, in a single raise, the entire financing ceiling it had promised to respect until 2030. Meanwhile the physical world sent two invoices: hackers set supermarket freezers to 50 °C across five countries, and Argentina offered its Patagonian wind, gas and land to anyone who wants to build compute, with no obligation to leave any of it behind.

Anthropic's revenue grew 12x in a year — and it still booked a $23.3 billion quarterly loss

Anthropic published an amended S-1 on the SEC website on Monday, November 2. Revenue for the quarter ended September 30 was $17 billion, up from $11.6 billion in the second quarter and $1.4 billion a year earlier — year-over-year growth of more than 1,000%. But sequential growth slowed sharply, to about 47% from about 144% the quarter before. Price reductions cut revenue by roughly $600 million between the two quarters.
Total operating expenses reached $17.5 billion against $10.6 billion in the second quarter, turning a $974 million operating profit into a $567 million operating loss, with $2 billion of that in non-cash charges tied to an employee donation matching program. Adjusted operating margin fell to 15% from 19%. The headline number is the $23.3 billion net loss, driven mostly by an accounting charge on the rising value of its convertible notes. The filing also discloses about $895.4 billion in committed spending, almost all of it compute and infrastructure, against $146.6 billion of liquidity as of September 30.

Why it mattersStrip out the accounting noise and one figure still matters more than the loss: $895.4 billion committed against $146.6 billion on hand. That gap is the business model of this entire cycle — sign for compute now, find the revenue later — and this is the first time a frontier lab has had to put it in a public filing where anyone can check the arithmetic. The deceleration from 144% to 47% in one quarter is the part the market will argue about for weeks. What it means for Latin America: that $600 million of revenue the company gave up to price cuts is your gain — frontier models are cheaper per token today than they were a quarter ago, and that is why AI products are suddenly viable from the region. But read the other side of it too: a company owing $895 billion in future compute will not keep subsidizing you forever. Build now at these prices, and design your product so the model underneath it can be swapped.

Marvell is paying $115 billion for Lumentum because AI now runs out of light before it runs out of chips

Marvell agreed on Monday, November 2 to acquire Lumentum, a maker of optical chips and networking equipment, in an all-stock deal valued at $115 billion including debt — with the equity portion around $80 billion according to a Wall Street Journal report cited by TechCrunch. Lumentum shareholders receive 4.5 Marvell shares for each share they hold. It is one of the largest acquisitions in tech history: for scale, Microsoft paid $68.7 billion for Activision Blizzard in 2022 and Dell $67 billion for EMC in 2016.
Nvidia is a major shareholder in both companies — it put $2 billion into Lumentum in March 2026 under a multi-year agreement to buy its advanced laser components, and $2 billion into Marvell later that same month. Marvell says the combination makes it "a leader in AI connectivity," spanning custom chips through to data center equipment. Closing is expected in the first half of calendar 2028, subject to shareholder and regulatory approval.

Why it mattersFor three years the scarce thing in AI was the accelerator. This deal says the scarce thing is now the cable between accelerators: when a training run spans a hundred thousand chips, the lasers and optical modules moving data between them become the ceiling on performance, and Marvell just paid more than Microsoft paid for all of Activision to own that layer. Note also that Nvidia funded both sides of this deal months before it happened — the supply chain is consolidating into a handful of overlapping balance sheets. What it means for Latin America: this is the layer that decides what your internet costs. The same optical components are what carry subsea cable and metro fiber capacity, and when hyperscalers absorb the supply, regional carriers wait longer and pay more for upgrades. If you are negotiating connectivity or colocation for 2027 and 2028, lock in bandwidth terms now rather than assuming prices keep falling.

Oracle raised $20 billion and spent its entire four-year AI financing promise in one move

Regulatory filings on Monday, November 2 show Oracle raising about $20 billion through a combination of bonds and share sales. The bond offering has six tranches maturing between 2029 and 2046 — one floating-rate and five fixed — with rates and prices not yet set, and the number of shares not yet determined. Proceeds go to general corporate purposes including capital spending, debt repayment, acquisitions, dividends and buybacks.
Notes payable and other borrowings stood at $125.34 billion as of August 31, against fiscal 2026 revenue of $67.36 billion. The company had pledged to raise no more than $20 billion in additional AI financing through fiscal 2030 — this single raise is that entire amount. Shares fell nearly 6% on Monday and are down about 33% this year; its existing 6.85% bonds due 2066 lost more than 1.2 cents on the dollar, per Tradeweb. CFO Hilary Maxson has said free cash flow should turn positive in the second half of fiscal 2028, and Oracle targets revenue above $225 billion by fiscal 2030.

Why it mattersOracle is the clearest test case for whether AI infrastructure can be financed with debt. It owes roughly $125 billion against $67 billion of annual revenue, capital spending peaks this year and next, and free cash flow is not expected to turn positive until 2028 — so every month between now and then has to be bridged with somebody else's money. The detail that should worry you is not the $20 billion; it is that the ceiling meant to reassure investors for four years was consumed in one filing. And the bond market noticed before the stock market did. What it means for Latin America: this is the quiet risk behind every cloud contract signed in the region. If the companies building the data centers are funding them with debt that reprices, the cost shows up later in list prices and in renewal terms. Keep your workloads portable, avoid multi-year commitments priced on today's promotional rates, and ask your provider where the capacity you are buying is actually being built.

Hackers set supermarket freezers to 50 °C across five countries, and the fix requires a screwdriver

A second wave of attacks on supermarket refrigeration began around midnight on Sunday night, November 1, according to Belgian agency Belga. Freezers were remotely set to 50 °C — about 122 °F — by activating each unit's built-in defrost mode and overriding its temperature and defrost-duration settings, the same method used in an attack on French stores weeks earlier. Belgium and Sweden were hit hardest, with reports also pointing to Ireland, Estonia and the Netherlands.
In Belgium the affected chains include Ahold Delhaize-owned stores, Delhaize, Carrefour and Spar, with refrigeration firm Sabcobel estimating at least 20 affected units nationwide; in Sweden, several ICA and Coop stores reported temperature spikes. No attacker has been identified. The uncomfortable part is the remedy: fixing the underlying vulnerability reportedly requires physical modifications to the units, so the interim recommendation is simply to move them onto a secured internet connection.

Why it mattersThis is the most legible industrial-security story of the year because the damage is food you can see rotting. The attack needed no malware and no credentials worth stealing — it used a feature the equipment shipped with, reachable because the controller was on the open internet. And there is no patch: the fix is a technician with a screwdriver visiting every unit, which means the exposure will persist for months after the headlines. Connected industrial hardware was sold on remote monitoring and nobody priced in remote sabotage. What it means for Latin America: the cold chain here runs on exactly the same remotely-managed controllers — supermarkets, butcher shops, pharmacies, vaccine storage, fish and produce exporters — and often on an open connection with a default password, because the installer set it up that way and nobody revisited it. Audit this week: list every refrigeration, HVAC or industrial controller reachable from the internet, change the default credentials, put them behind a VPN, and set a temperature alarm that calls a human rather than just logging the event.

Argentina is offering Patagonia to AI data centers with a 15% tax rate and no rules attached

Reporting published November 2 describes Argentina courting AI data centers with deregulation. Polish renewable energy firm Green Capital plans a $3 billion data center in Patagonia to lease to big tech, with no customer signed yet; 14 landowners have leased it a combined 130,000 hectares — about 321,000 acres — near Trelew in Chubut province, with construction not expected until around 2028. Power would come from Vaca Muerta shale gas plus the company's own wind, solar and gas generation. At projected peak demand the campus would draw roughly 2.7 times Buenos Aires' entire 2025 annual electricity consumption. How much water it needs has not been disclosed.
President Javier Milei has proposed cutting the corporate income tax rate from 25% to 15% for projects of at least $1 billion in AI and other emerging industries — the Senate has not approved it — and has said Argentina will not regulate AI. Critically, the incentives do not require companies to reserve any computing capacity for local use or to disclose energy and water consumption. Crusoe, the developer behind OpenAI's Texas campus, has sought talks with Argentine energy officials, and Milei's meeting logs include OpenAI, Amazon, Google and Oracle. A researcher at Observatorio Petrolero Sur argues the capacity would simply be exported, while a computer science professor at the National University of Córdoba warns the country could become a testing ground; a "Trelew Free of Data Centres" movement has formed locally.

Why it mattersThis is the most important Latin American tech story of the day, and it is a warning about the shape of the deal, not about data centers themselves. The region's real asset in this cycle is cheap energy and cool land — yesterday Google's own energy chief confirmed terrestrial compute beats orbital by a factor of a hundred — so demand is coming here regardless. The question is what gets written into the contract. A 15% tax rate with no obligation to reserve local compute, no disclosure of water or power draw, and a campus sized at 2.7 times what Buenos Aires uses in a year is not an industrial policy; it is a lease on the cheapest input and nothing more. What it means for Latin America: the competition between Argentina, Brazil, Chile and Mexico should be a competition over terms, not a race to the lowest tax rate, because tax rates are the one thing a capital-intensive project barely notices and the one thing citizens lose permanently. The terms worth demanding are specific and perfectly normal elsewhere: a reserved percentage of compute for local universities and companies, published energy and water figures, grid-flexibility commitments, and a price floor on the electricity sold. If you work in policy, procurement or press in the region, that list is your checklist for the next announcement — and it will arrive soon.

The big picture

Put Monday's filings side by side and the story is not that AI got smarter. It is that AI's bill arrived, and each player chose a different way to pay it. Anthropic chose disclosure: $895.4 billion committed against $146.6 billion of liquidity, printed in a public document where the gap is now everyone's to judge. Oracle chose leverage, and burned a promise meant to last four years in a single $20 billion raise while its own bondholders marked it down. Marvell chose to buy the bottleneck outright, $115 billion for the lasers that move data between chips, in a deal Nvidia had quietly funded from both ends. And the physical world sent the other two invoices on the same day: supermarket freezers in five countries turned into a demonstration that connected industrial equipment has no patch, and Patagonia turned into a demonstration of what energy is worth to someone who needs it badly. Those last two are the ones that touch the region directly, and they point the same direction. The things that are actually scarce in this cycle are power, land, optical capacity and secured industrial networks — all four of which Latin America has or needs to protect — and the things that are abundant are capital and promises. If you build from here, price your hardware and connectivity before the next quarter does it for you, audit the controllers you forgot were on the internet, and when someone arrives offering jobs in exchange for megawatts, negotiate for compute and transparency instead of a lower tax rate.

Go deeper

Sources & further reading

Topics:AI Chips Big Tech Regulation