Today's 12 Stories — Tuesday, November 3, 2026

Funding & Investment Story 1 of 12

Anthropic's Amended IPO Filing Shows $17.0 Billion in Quarterly Revenue and $895.4 Billion in Compute Commitments

Anthropic gave public market investors their first full look at its third quarter on Monday, filing an amended S-1 registration statement with the Securities and Exchange Commission. The numbers are enormous in both directions, and they will frame every conversation about the company's planned listing on the Nasdaq Global Select Market under the ticker ANT.

According to the filing, Anthropic recorded revenue of $17.0 billion for the three months ended September 30, 2026. The company says that figure is up 1,127% year over year, while its compute and infrastructure expenses grew 382% over the same period. Anthropic presents that gap as evidence of compute efficiency: revenue is compounding much faster than the cost of the chips and data centers behind it.

The profit picture is more complicated. The filing reports an operating loss of $567 million for the quarter, a swing from operating income of $974 million in the three months ended June 30. Anthropic attributes the reversal primarily to $5.0 billion of compute bought under short term commitments, plus $2.0 billion of non cash charges tied to its donation matching program. Below the operating line, the filing's financial tables show a net loss of $23.291 billion for the quarter.

The figure that will draw the most scrutiny is the size of the bet Anthropic has already placed. The filing says that, as of and after September 30, the company had entered into purchase commitments and lease agreements with an aggregate committed spend of approximately $895.4 billion, substantially all of it for compute and infrastructure. Approximately $447.5 billion of that is expected to be incurred before the end of 2030.

The company is also far larger as an organization than it was two years ago. Its filing puts headcount at approximately 5,300 as of September 30, up from approximately 940 at the end of 2024.

For executives, the filing reads less like a growth story and more like a capital allocation thesis. Anthropic is telling investors that demand for frontier models is durable enough to justify commitments on the scale of national infrastructure budgets, and that the economics improve as the company shifts from expensive spot capacity to long term contracts. The quarter's operating loss shows what happens when demand outruns planned supply.

For buyers of AI services, the practical read is about supplier stability and pricing power. A vendor carrying commitments of this size has every incentive to lock in large enterprise contracts and to keep usage growing. Procurement teams negotiating multi year agreements should expect that dynamic to shape terms, and should watch how public market investors price the risk once the offering is underway.

AI Safety Story 2 of 12

Anthropic Tells Investors It May Soon Hit Its Own Threshold for AI Accelerated Research

Buried in the risk factors of Anthropic's amended S-1, filed Monday, is a disclosure with implications well beyond the IPO. The company told investors that it may be close to a point it has long described as one of the most consequential in AI development: models that dramatically speed up the work of building the next generation of models.

The filing states that Anthropic's internal estimates and expectations regarding future models, rather than its current models, indicate that it may imminently meet its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI." The company adds that it has not designated any of its current models as exceeding this threshold and that it continues to assess the underlying data.

The context makes the warning concrete. Anthropic's filing says that over 90% of the contributions to its codebase in September 2026 were written by Claude. In the same risk factor, the company cautions that its growing use of AI to accelerate its own research could cause model capabilities to advance faster than its ability to develop and implement matching safety, alignment and security measures, and that it may not recognize a significant acceleration until after it has happened.

Anthropic also drew a line. The filing states that it does not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible. Accordingly, the company says it is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development.

The disclosure arrives in a document that is both a legal obligation and a marketing instrument, and skeptics will read it with that in mind. But a securities filing carries liability that a blog post does not. A company does not casually tell prospective shareholders that its own safety framework may soon be triggered.

For boards and risk committees, the significance is twofold. First, this is a frontier lab stating in a regulated document that the feedback loop between AI and AI research is tightening on a measurable timeline. Policymakers who have treated that scenario as speculative now have a filing to point to, and the disclosure is likely to surface in legislative debates on frontier model oversight.

Second, it raises a governance question for every enterprise that relies on frontier models. If capability gains can arrive faster than the vendor itself can evaluate them, customers need contractual visibility into how model updates are tested and rolled out. Change management clauses, evaluation commitments and the right to pin model versions deserve a fresh look in any major AI agreement.

Industry Dynamics Story 3 of 12

Marvell Agrees to Buy Lumentum in an All Stock Deal Valuing the Optical Maker at $115 Billion

Marvell Technology announced on Monday that it will acquire Lumentum in an all stock transaction, a deal that puts optical connectivity at the center of the AI infrastructure race. Under the definitive agreement, Lumentum stockholders will receive a fixed 4.500 shares of Marvell common stock for each Lumentum share they own.

The companies said the transaction implies an enterprise value of $115 billion for Lumentum, or approximately $1,132 per Lumentum share. That represents a premium of approximately 23.9% based on unaffected closing prices on October 28, 2026, the last trading day before market speculation about a deal began. After closing, current Marvell stockholders are expected to own approximately 67% of the combined company and Lumentum stockholders approximately 33% on a fully diluted basis.

The strategic logic is about physics as much as finance. Marvell argues that bandwidth demands in AI clusters, and the limits of copper wiring, are pushing the industry toward optical technologies, and that the next generation of systems will require much tighter design of electrical and optical components together. The deal pairs Marvell's connectivity and custom silicon business with Lumentum's lasers and optical components.

Marvell says the combination expands its total addressable market to $550 billion and provides line of sight to more than $1 billion in annual run rate cost synergies. The company also expects the deal to be accretive to its non GAAP earnings per share in the first year after closing.

"AI infrastructure is undergoing a fundamental architectural shift toward optical connectivity, creating an opportunity that comes along once in a generation," said Matt Murphy, Marvell's chairman and CEO.

The transaction is expected to close in the first half of calendar year 2028, subject to stockholder and regulatory approvals. Murphy will lead the combined company, and three Lumentum directors, including Lumentum CEO Michael Hurlston, will join Marvell's board.

For technology leaders, the deal signals where the bottleneck in AI data centers is moving. Compute gets the headlines, but moving data between accelerators is increasingly what limits cluster performance and energy efficiency. A vertically integrated supplier that can design the electrical and optical signal chain together could shorten the path from chip to deployed system.

The long timeline matters too. A close in 2028 means more than a year of regulatory review across multiple jurisdictions for a deal touching strategically sensitive optical and semiconductor manufacturing. Buyers of networking hardware should expect both companies to keep selling independently until then, and should watch whether regulators attach conditions on supply to rival platforms.

Enterprise AI Story 4 of 12

Palantir's U.S. Commercial Revenue Jumps 148% and Overtakes Its Government Business for the First Time

Palantir reported third quarter results on Monday that marked a milestone in the company's history: its U.S. commercial business is now bigger than its U.S. government business.

Total revenue rose 96% year over year to $2.313 billion, the first quarter in which Palantir's revenue exceeded $2 billion. U.S. commercial revenue grew 148% year over year to $984 million, edging past U.S. government revenue of $953 million, which itself rose 96%. For a company built on defense and intelligence contracts, the crossover is a structural shift in its customer base.

Profitability kept pace with growth. Palantir reported GAAP net income of $1.128 billion for the quarter and a Rule of 40 score of 158%, the metric software investors use to balance growth against margins.

The company also raised its outlook. Palantir now expects full year 2026 revenue of between $8.530 billion and $8.534 billion, and U.S. commercial revenue of more than $3.559 billion, which would represent growth of at least 143%.

Chief executive Alex Karp framed the results around what he calls sovereign AI. "The Sovereign AI movement is imposing the most dramatic dichotomy between haves and have nots that I have witnessed in 30 years of business," Karp said in the earnings release, arguing that every institution is being pushed into one of those two categories.

Beneath the rhetoric is a concrete enterprise pattern. Palantir's pitch is that large organizations get value from AI not by buying a model but by connecting models to their operational data and workflows, which is what its Foundry, Ontology and AIP products are designed to do. The commercial growth rate suggests that pitch is landing with American companies that have moved past pilot projects and want production deployments.

For executives evaluating AI platforms, the quarter offers two lessons. First, the market is rewarding vendors that sell integration and governance around models, not just access to them. The hard part of enterprise AI remains data plumbing, permissions and change management, and buyers are paying for help with it.

Second, the crossover from government to commercial revenue reduces Palantir's dependence on federal budgets, but it also places the company in direct competition with the hyperscalers and enterprise software incumbents for the same corporate budgets. Expect pricing pressure and bundling offers from rivals as they respond, which strengthens the negotiating position of large buyers in upcoming renewals.

AI Infrastructure Story 5 of 12

Oracle Files to Raise $20 Billion Through New Stock and Senior Notes

Oracle moved to raise $20 billion on Monday, filing preliminary prospectus supplements for a $10 billion offering of common stock and a concurrent $10 billion offering of senior notes. The filings are the latest sign of how much capital the AI data center buildout is pulling from public markets, and of how the largest builders are now tapping equity as well as debt.

According to the equity prospectus, the closing of the stock offering is not conditioned on the notes offering, and vice versa. The stock will rank junior to the notes. Oracle's filing notes that its shares last traded at $130.67 on the New York Stock Exchange on October 30, 2026.

Oracle says the net proceeds will be used for general corporate purposes, which may include capital expenditures, repayment of indebtedness, future investments or acquisitions, and dividends or share repurchases. The prospectus does not earmark the money for a specific project, but its forward looking statements leave little doubt about where the pressure lies. Among them is Oracle's expectation that customer demand will require continued growth in capital expenditures to increase existing data center capacity and add new locations, and its ability to secure reliable and cost effective power for those data centers.

The decision to sell new shares alongside bonds is notable. Debt has been the default way to finance data center construction, because it does not dilute existing shareholders. Issuing equity at this scale suggests that Oracle, like other heavy spenders, is managing its balance sheet and credit profile as commitments mount, rather than leaning on borrowing alone.

The filing lands in a market that has grown more skeptical of AI spending. Investors have spent recent weeks questioning whether the revenue from AI services will arrive fast enough to cover the cost of the infrastructure being built for it, and companies that need outside capital to keep building face that question most directly.

For enterprise customers, the practical takeaway concerns capacity and contracts. Oracle's cloud infrastructure business depends on adding data center capacity quickly, and the raise is designed to fund that growth. Customers with large cloud commitments benefit if the capacity arrives on schedule. But a provider financing expansion through both equity and debt is also one whose costs, and pricing, are tied to capital markets conditions.

Chief financial officers should treat the filing as a data point in vendor risk assessments: the AI cloud providers are now competing not only for customers but for investors, and the terms they win in capital markets will shape the terms they can offer.

Policy & Regulation Story 6 of 12

The New York Times and Three Other Papers Sue Brave, Alleging It Sold Scraped Journalism to AI Companies

A new front opened in the fight over AI training data on Monday, and this time the defendant is not a model developer. The New York Times sued Brave Software, the company behind the Brave web browser, alleging unlawful scraping of its content for AI and a browser feature that lets users bypass paywalls.

A separate complaint filed by the New York Daily News, the Chicago Tribune and the Orlando Sentinel goes further into the business model. According to that complaint, Brave has monetized its index of scraped websites in part by selling AI companies paid access to "real time online content from the Brave Index." The AI companies then use that index to train their models, to answer specific user queries, or both. The complaint names Anthropic, Perplexity, Mistral, xAI and Cohere among the companies that allegedly received the content.

The publishers argue that Brave never sought permission to copy their work or to pass it on to users and AI companies, and that it never paid for it. The allegations have not been tested in court.

The cases matter because they target the supply chain rather than the end user of the data. Most copyright litigation against AI has focused on the companies that train models. These suits ask whether an intermediary that crawls the web and resells real time access to that content is itself liable, and whether a buyer of such a feed can claim distance from how it was gathered.

For enterprises building AI products, the immediate lesson is about provenance. Retrieval systems increasingly depend on third party search and web index APIs to ground answers in current information. If courts treat the resale of scraped content as infringement, the legal exposure could travel downstream to anyone whose product depends on those feeds.

General counsels should ask their vendors a direct question now: where does your retrieval data come from, and what rights do you hold to it? Contracts for search APIs and data feeds should carry indemnities and clear representations about licensing. This case suggests that a license in name only may not be enough.

Policy & Regulation Story 7 of 12

Putin Orders Private Data Centers to Reserve 10% of Capacity for the State

Russian President Vladimir Putin signed a decree on Monday requiring Russian owners of non state data centers to reserve 10% of their capacity to help keep the country's IT infrastructure running, according to Reuters reporting carried by The Moscow Times. The measure gives the state a claim on privately operated computing capacity at a moment when that capacity has become a military target.

Under the decree as reported by Reuters, a special government commission will be able to order the release of the reserved capacity, and operators will be required to provide access within 24 hours of such a decision. Putin instructed the Digital Development Ministry to conduct an inventory of available capacity at privately owned data centers across Russia. The decree also allows for the temporary use of data centers located abroad, subject to approval by the FSB security service.

The backdrop is a string of attacks. Data centers operated by Yandex and Sberbank were struck by Ukrainian drones in October, according to the same reporting. The decree effectively treats commercial data centers as strategic reserve infrastructure, similar to fuel or grain stockpiles, that the state can draw on in an emergency.

The move fits a broader global pattern, even if the circumstances are extreme. Governments everywhere are reclassifying computing capacity from a commercial commodity to a matter of national resilience. Sovereign AI programs, data localization rules and emergency powers over critical infrastructure all reflect the same instinct: compute is now too important to leave entirely to the market.

For multinational companies, the Russian decree is a reminder that a data center's legal jurisdiction can override its commercial contract. A provider subject to a mandatory reserve or seizure order may be unable to honor capacity guarantees to its own customers, whatever its service level agreement says. Few organizations still run workloads in Russia, but the principle applies wherever governments hold broad emergency powers over infrastructure.

Risk teams should review where critical workloads physically run and which governments can compel access to those facilities. Contingency plans that assume a cloud region will remain available under all conditions deserve a second look, particularly for operations in regions exposed to conflict or to governments with expansive emergency authority.

The decree also signals that physical security is becoming a design requirement for AI infrastructure, not an afterthought. As data centers grow more central to economies and militaries alike, their protection, redundancy and geographic distribution will increasingly be shaped by state priorities rather than purely by cost and latency.

AI Infrastructure Story 8 of 12

OpenAI Pitches a Georgia Data Center as a Grid Asset That Can Give Back Up to 1,000 Megawatts

OpenAI announced on Monday that it will pilot what it calls a Grid Protectors Program at a planned data center in Effingham County, Georgia. The company says the site will be able to free up to 1,000 megawatts of power during extreme weather or emergencies, handing electricity back to the grid when local communities need it most.

The announcement is as much about public opinion as about engineering. Local resistance to large data centers has become a political issue in many communities, driven by worries about power prices, water and pollution. OpenAI's response is to recast its facilities as a resource that utilities can call on in a crisis rather than a permanent drain on the system.

The underlying argument is about how grids are paid for. A facility that draws power steadily all year helps spread a utility's fixed costs across more electricity sales, and if it can also cut its demand during the few hours when the system is most stressed, it avoids forcing the utility to build capacity just for those peaks.

The details of how much computing work would be paused, on what notice, and how often, will determine whether the commitment is meaningful in practice. Those terms have not been laid out in public, and they are the part utility regulators will want to see.

Demand flexibility is not a new idea. Industrial users such as aluminum smelters and crypto miners have long been paid to curtail load during peaks. What is new is the scale, and the claim that AI training and inference workloads can be scheduled around the grid's needs without breaking service commitments to customers.

For executives, the program points to a shift in how AI infrastructure gets permitted. Access to power is now the binding constraint on new data centers, and communities hold real leverage through zoning, utility commissions and state legislatures. Companies that can demonstrate a tangible benefit to local ratepayers are likely to move through approvals faster.

Enterprises planning their own large compute footprints, or negotiating with colocation providers, should expect demand response terms to become standard. That has an operational consequence: workloads that can tolerate interruption will be cheaper to run than those that cannot. Architects should start classifying AI jobs by their tolerance for pauses, because the energy market is about to price that difference.

AI Business Models Story 9 of 12

SUSE Cuts About 10% of Its Staff and Names AI as a Reason, as Squarespace Lays Off 19%

Two software companies announced significant job cuts on Monday, and one of them said plainly that AI was part of the reason.

SUSE chief executive Dirk-Peter van Leeuwen told staff in a memo that the open source software company will cut about 10% of its roughly 1,900 employees, The Register reported. According to the report, the memo cited automation and AI as forces changing how software companies operate and reducing the need for some existing roles. Affected staff were to be notified by November 4. SUSE is owned by the private equity firm EQT.

Separately, Squarespace is reducing its global workforce by approximately 19%, around 300 employees, a company representative confirmed to TechCrunch. Squarespace did not attribute the layoffs to AI, but the website building market it serves has been reshaped by AI tools that let users generate sites from plain language prompts.

The two announcements illustrate a distinction executives should keep clear. Some companies are explicitly reducing roles because AI tooling changes how work gets done inside the business. Others are restructuring because AI is changing the competitive landscape for their products. Both lead to fewer jobs, but they call for different responses.

The SUSE memo is notable for its candor. Many companies have avoided naming AI as a driver of layoffs, wary of the reputational and regulatory attention it invites. A chief executive stating in writing that AI tooling has reduced the need for some roles is a sign that the framing is becoming normalized, at least inside software firms where engineers adopted coding assistants early.

For leaders weighing similar decisions, the lesson is about sequencing and evidence. Cutting roles on the expectation of productivity gains, before those gains are measured, risks losing institutional knowledge that AI tools cannot replace. The more defensible path is to document where automation has actually reduced workload, redeploy people where possible, and be transparent about the rationale.

Boards should also expect scrutiny. Labor regulators, works councils in Europe and employees themselves are paying closer attention to how AI is cited in workforce decisions. Clear documentation of how a decision was reached, and of what support affected employees receive, is becoming part of responsible AI governance rather than a separate HR matter.

Enterprise AI Story 10 of 12

Goldman Sachs Pulls Advanced AI Coding Tools From Traders and Bankers

Goldman Sachs has limited advanced AI coding tools such as Anthropic's Claude Code largely to its engineers, cutting off most traders and bankers, according to Bloomberg reporting on Monday that cited internal guidance updated in October. The move reverses months in which non technical staff used the tools to build their own software, and it offers a rare look at how a major bank is drawing the boundaries of AI use.

Traders and bankers had spent months using the tools to monitor trading flows and create content, according to the reporting. As recently as July, Goldman chief information officer Marco Argenti had praised vibe coding by non technical staff in an interview.

A Goldman spokesperson, Nick Carcaterra, attributed the change to factors including productivity and risk management.

The decision cuts against one of the most hyped promises of agentic coding tools: that anyone in an organization can become a software builder. In a heavily regulated institution, software written by non engineers creates obvious questions. Who reviews the code? Who maintains it when the author changes roles? How does it fit into change control, audit trails and model risk frameworks that regulators expect banks to follow?

Goldman's answer, at least for now, is that the governance cost outweighs the productivity benefit outside the engineering organization. That is a meaningful data point for every enterprise weighing broad rollouts of coding agents.

For chief information officers, the lesson is not that citizen development is a mistake, but that it needs a defined path to production. Prototypes built by business users can be valuable, but they become a liability when they quietly turn into critical tools without review. Organizations that want the benefit without the risk should define sandboxes, mandatory engineering review for anything that touches production data, and clear ownership rules.

Chief risk officers should also note the pattern of rapid expansion followed by retrenchment. Pilot programs that put powerful tools in many hands can generate shadow software faster than oversight can track it. Inventorying what was built during the experimental phase, and deciding what to retire or formalize, is the unglamorous work that should accompany any change in access policy.

AI Models Story 11 of 12

Meituan Releases LongCat-2.5, a 1.6 Trillion Parameter Multimodal Model Built for Long Agent Tasks

Meituan, the Chinese food delivery and local services giant, released LongCat-2.5 on Monday, a large multimodal model aimed squarely at the long running, multi step work that AI agents are expected to take on. The model has 1.6 trillion total parameters, with about 48 billion active for each token it processes. Because only a small fraction of the network runs on any given token, the cost of serving it sits far below what the headline size suggests.

The training approach is the more interesting story. According to TechNode, Meituan used large scale reinforcement learning aimed at complex codebases and complete software projects, and it converts failures encountered by real agents into reproducible training tasks, a feedback loop that turns production mistakes into curriculum. LongCat-2.5 is available through Meituan's online chat service and its API platform.

Meituan's presence at this scale is itself the signal. A company best known for delivering meals is now shipping a model in the trillion parameter class, reflecting how deeply Chinese technology groups have committed to building their own frontier capabilities rather than depending on a small number of labs.

The emphasis on long horizon agent work also tracks the direction of the whole industry. The benchmark battles of recent years focused on single answers to single questions. The competition now is over whether a model can stay on task across hours of work, recover from errors and finish a project. Training on real agent failures is one of the clearest bets on that shift.

For technology leaders, the release widens the field of capable models. Enterprises with operations in China, or those evaluating open weight options for cost or sovereignty reasons, have another serious candidate to test. The mixture of experts architecture, with its relatively small active parameter count, makes the inference economics worth examining.

The usual cautions apply. Organizations should run their own evaluations on representative tasks rather than relying on vendor claims, and should assess data handling, licensing and regulatory exposure before deploying any model across borders. The practical question is not which model tops a leaderboard, but which one completes your workflows reliably at an acceptable cost.

AI Research Story 12 of 12

ElevenLabs Data Shows Voice Agents Beat Chat by 31.7% on Longer Customer Conversations

ElevenLabs published research on Monday comparing voice agents and text chat in real customer service deployments, and the findings offer a practical guide for companies deciding where to invest. The short version: chat wins quick exchanges, but voice pulls ahead as conversations get longer and more complex.

The company analyzed agents that handled at least 50 voice and 50 chat conversations between May and July 2026, judging each against success criteria the business itself had defined. To avoid skewing the comparison, each agent's voice performance was compared against its own chat performance rather than pooling all conversations.

Using conversation length as a stand in for complexity, a clear pattern emerged. In conversations under one minute, chat outperformed voice by about 9 percentage points. Past two minutes, voice success rates were 10.9 percentage points higher, a 31.7% relative increase. Among successful conversations over two minutes, voice averaged 4.6 minutes compared with 7 minutes for chat.

Speed of response turned out to matter as much as the channel. ElevenLabs said that across 315 agent conversations that switched to a faster LLM, a 16% reduction in average end to end latency correlated with a 26% lift in success rate. The company also reported a user study in which 98 participants worked through a simulated canceled flight with both a voice agent and a chat agent; ElevenLabs said 85% preferred voice.

The usual caveat applies: this is a vendor publishing research about a category it sells, and conversation length is an imperfect proxy for complexity, especially in chat, where customers may pause between messages. But the methodology of comparing each deployment against itself is sound, and the direction of the findings matches what contact center leaders have long observed with human agents.

For customer experience leaders, the takeaway is to stop treating voice and chat as interchangeable channels. Routine, transactional requests such as order status or password resets fit chat well. Stressful, multi step problems, such as rebooking travel, disputing a charge or untangling a billing error, appear better suited to voice.

The latency finding deserves particular attention. Many companies select models for their reasoning quality and treat response time as secondary. For voice, that ordering may be backwards: a slower, smarter model can lose callers who assume the line is dead before it answers. Teams building voice agents should measure end to end latency as a primary success metric, not a technical footnote.