Today's 12 Stories — Wednesday, November 4, 2026

AI Infrastructure Story 1 of 12

AMD's Data Center Revenue Nearly Doubles to $8.5 Billion as MI450 GPUs Ship

AMD reported third quarter 2026 revenue of $13.4 billion, and the number that matters most to anyone planning AI capacity sat inside it. The company's Data Center segment brought in a record $8.5 billion, up 97% year over year, as the first Instinct MI450 Series GPUs reached customers alongside strong demand for EPYC server processors.

The company named the buyers. In its results, AMD said Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure are running models and workloads on its first Helios systems. That list reads like a roll call of the largest AI compute spenders in the world, and it is the clearest evidence yet that the frontier labs are treating AMD as a real second source rather than a negotiating tactic.

The guidance is even more striking than the quarter. AMD expects fourth quarter 2026 revenue of approximately $17 billion, plus or minus 3%, which would be a step change from the quarter just reported. Chief executive Lisa Su said she expects growth to accelerate further in the fourth quarter as the company scales Helios and Venice production.

Not every line moved in the same direction. Gaming revenue fell to $773 million, down 40% year over year, a reminder that AMD's consumer business and its AI business now live in very different markets. The company's center of gravity has shifted decisively toward the data center, and the income statement now says so plainly.

For executives, the practical consequence is leverage. A credible second supplier of high end accelerators changes the economics of every large compute contract. Buyers that once had to accept a single vendor's allocation, pricing and software roadmap can now credibly split orders, and the names AMD listed suggest the largest buyers are already doing exactly that.

The risk sits in execution. Rack scale systems are far harder to deliver than individual chips, because they bundle networking, cooling, power and software into a single product that must work on day one. AMD's fourth quarter guidance assumes that ramp goes smoothly. If it does, the AI accelerator market will look meaningfully less concentrated in 2027 than it does today. If it stumbles, customers will have learned that a second source is only as good as its delivery schedule.

Either way, procurement teams should be asking their cloud providers a direct question this quarter: which of our workloads can run on AMD hardware, and what would it cost if they did.

AI Infrastructure Story 2 of 12

Supermicro Posts $19.9 Billion Quarter and Raises Its Fiscal 2027 Outlook to as Much as $82 Billion

Supermicro reported first quarter fiscal 2027 revenue of $19.9 billion, and on its earnings call the company said that figure was up 298% year over year. The server maker then raised its full year outlook, telling investors it now expects fiscal 2027 revenue of $72 billion to $82 billion, up from a prior range of $65 billion to $72 billion.

The company guided second quarter fiscal 2027 revenue of $19.5 billion to $21.5 billion, according to its earnings call, implying that the extraordinary first quarter was not a one time spike but a new baseline. Founder and chief executive Charles Liang described the year as off to a record start.

The physical numbers are as telling as the financial ones. On the call, Supermicro said it is on track to produce more than 8,000 racks per month by December, including more than 4,000 direct liquid cooling racks. That second figure deserves attention. Liquid cooling has moved from a specialist option to a mainstream requirement, because the newest AI accelerators generate more heat than traditional air cooled data centers can remove. A manufacturer planning to build thousands of liquid cooled racks a month is a signal about what the next generation of AI facilities will look like inside.

Supermicro occupies an unusual position in the AI supply chain. It does not design the accelerators that dominate headlines, but it assembles the servers and full racks that turn those chips into usable compute. When its revenue surges, it is a direct read on how quickly cloud providers, newer GPU cloud companies, enterprises and sovereign buyers are converting purchase orders into installed capacity.

That makes this quarter a useful counterweight to the ongoing debate over whether AI infrastructure spending has peaked. A company that sells the boxes is raising its forecast, not cutting it.

The caution for investors and buyers is the familiar one for this business. Server assembly carries thin margins, and Supermicro's results depend heavily on a small number of very large customers and on steady supply of chips it does not control. A rapid buildout can reverse just as quickly if a major buyer pauses.

For enterprise leaders, the takeaway is about lead times and facilities. If you are planning on premises AI capacity in 2027, the industry is now building for liquid cooled density by default. Facilities that cannot support it will narrow your hardware options, and that conversation with your real estate and operations teams should start now, not after the purchase order.

Policy & Regulation Story 3 of 12

China's Economic Planner Says Beijing Will Write AI Safety Rules for the Full Life Cycle of the Technology

China's state economic planner has signaled one of the broadest attempts yet by any government to set guardrails around increasingly powerful AI. At a briefing in Beijing on November 3, Wang Ruomeng of the National Development and Reform Commission said Beijing will establish safety standards and ethical guidelines covering the full life cycle of AI, from development through deployment.

The plan, as Wang described it, reaches well beyond a single rulebook. Beijing intends to accelerate legislation on the core enablers of AI, namely data, computing power and algorithms, and to regulate key application scenarios. The commission said AI developers and operators should strengthen risk management and disclose safety risks, incidents and vulnerabilities. It also said it will create sector specific regulatory sandboxes where companies can test new applications, and that it will monitor AI's impact on employment while providing safety net protections for affected workers.

The announcement arrives against a backdrop of rising international alarm. Bloomberg, which reported on the briefing, tied it to growing global concern following warnings from Silicon Valley leaders about existential risk and a string of recent incidents in which AI systems behaved unexpectedly or enhanced hacking attempts.

The framing matters for multinational companies. A life cycle approach implies rules that follow an AI system from the data it is trained on to the way it is deployed, a far more comprehensive ambition than regulating one application at a time.

Beijing is also trying to have it both ways, and saying so. Wang said China would keep encouraging open AI development through its current five year plan and would ease compliance burdens through a transition period and official guidance. The message to domestic champions is that the state wants safety rules without slowing the open model ecosystem that has become a point of national pride.

For executives with operations or partners in China, three details deserve tracking. First, disclosure of incidents and vulnerabilities would create reporting obligations that do not exist in most markets today. Second, the sandboxes may offer earlier paths to deploy new applications in regulated sectors. Third, the commission said it wants deeper international cooperation on safety governance, including coordinated vulnerability disclosure, which could eventually shape how global companies report AI flaws across jurisdictions.

The details are still to come. But the direction is now clear: the world's two AI superpowers are no longer only competing on capability. They are beginning to compete on who writes the safety rules.

Industry Dynamics Story 4 of 12

Canva Restructures 1,271 Jobs as Its Growth Slows for the First Time in a Decade

Canva, the online design platform, announced a restructuring affecting 1,271 employees, a share reported at about 22 percent of its workforce. The changes include job cuts, role changes and relocation requests.

Chief executive Melanie Perkins did not soften the message. In a staff memo, she wrote, "This is without doubt the hardest moment in our thirteen-year history." She told employees that Canva's growth had slowed for the first time in a decade and that the company's structure had made it harder for teams to take ownership across departments.

The timing is not subtle. Canva built its business by making professional looking design accessible to people who are not designers. Generative image tools from the largest AI labs now promise something similar. According to coverage of the cuts, the restructuring comes as Canva faces growing competition from AI image generation tools, including those from OpenAI and Google, although Perkins herself pointed to slowing growth and organizational structure.

Canva is trying to treat departing staff generously. The company said affected employees will receive at least six months' salary and 12 months of continued healthcare coverage where Canva already provides it, along with career support. It is also reshaping how the people who remain will work. From January 2027, most hybrid staff will be expected in the office two to three days a week, and Sydney will become the main base for more product teams.

The broader lesson extends well beyond design software. Canva is a case study in a risk many software companies have quietly acknowledged: a product built on making a hard task easy is exposed when AI makes that task easier still. Feature parity is no longer the main threat. The threat is that the task itself gets absorbed into a general purpose assistant.

For leaders running software businesses, the questions Canva now faces are worth asking early. Which of our core workflows could a frontier model perform end to end? Where do we hold proprietary data, distribution or trust that a general model cannot replicate? And is our organization structured to ship AI native products quickly, or, as Perkins described, are our teams too fragmented to own outcomes?

Canva still has a large user base and a strong brand. Its restructuring is less an obituary than an early warning. The companies that answer those questions before growth slows will have more options than those that answer them afterward.

AI Safety Story 5 of 12

CrowdStrike Moves to Buy Fiddler and Dreadnode and Ships a Kill Switch for Rogue AI Agents

CrowdStrike made one of the clearest bets yet that securing AI agents will become its own category of enterprise spending. The company signed a definitive agreement to acquire Fiddler AI, a specialist in monitoring and explaining what AI agents do, and agreed to acquire Dreadnode, which tests AI models, applications and agents for weaknesses before they are deployed.

Fiddler's technology will be sold inside CrowdStrike's Falcon platform as Falcon Observe. Chief executive George Kurtz described it this way: "Fiddler is the flight recorder for every agent: the observability to explain exactly what an agent did and why." Dreadnode comes at the problem from the other side, running more than 100 known AI attack strategies, including jailbreaks and prompt injection, to find failures before attackers do.

CrowdStrike did not disclose the purchase prices. It said payment will be mostly cash, with some stock subject to vesting conditions, and it expects both deals to close in its fourth quarter of fiscal 2027.

Alongside the acquisitions, CrowdStrike launched a feature with a deliberately blunt name. The AI Kill Switch, part of its Falcon Guardian product, lets an authorized administrator halt a rogue AI agent where it runs, prevent it from relaunching and cut its access to the corporate network. Every action is logged.

Taken together, the moves describe a full life cycle for agent security. Dreadnode tests an agent before deployment. Fiddler watches what it does in production and explains its reasoning. The kill switch stops it when something goes wrong. That is the same pattern security teams already use for software and endpoints, applied to a new kind of actor that can take actions on its own.

The strategic logic is straightforward. As companies give agents access to email, code repositories, financial systems and customer data, every agent becomes a potential insider with credentials. Traditional security tools were built to watch people and programs, not software that reasons, plans and changes course mid task. Security vendors that can see and control agent behavior will own a budget line that barely existed two years ago.

For chief information security officers, the practical question is inventory. Most organizations cannot yet list every agent running in their environment, what it can access, or who approved it. Buying an observability product does not solve that on its own. But CrowdStrike's moves are a strong signal that boards and auditors will soon expect answers, and that an agent with no off switch will look like an unmanaged risk.

Generative AI Story 6 of 12

Apple Opens Siri AI to Every App Developer, but Not Yet in the EU

Apple has opened its new Siri AI to outside developers. Starting with the iOS 27.3, iPadOS 27.3, macOS 27.3 and visionOS 27.3 betas, a new developer preview lets any developer integrate an app with Siri AI using App Intents, the framework apps use to bring their content and actions into Siri AI.

The practical promise is that people can ask for help in their own words and Siri AI will carry out a custom action an app has built, from checking a delivery to booking a reservation or placing an order. Apple says apps can tap Siri AI's newer capabilities, including personal context, onscreen awareness and natural language understanding.

Apple also introduced what it calls skills, which let developers teach Siri AI to combine several actions into a single experience. That is the most consequential detail in the announcement. A single action is a voice shortcut. A chain of actions across steps is an agent, and Apple is now inviting every app on its platforms to become a component in agentic workflows that Siri orchestrates.

The timeline is measured. Apple said this preview provides early access before first shipping to customers in a few months, with App Store submission opening in the spring. Developers have a window to build and test, and Apple has time to set the rules, which it said it will detail later in the form of terms and usage policies.

One region is conspicuously absent. Apple said Siri AI is currently unavailable in the EU in iOS, iPadOS and watchOS, and that it is working with the EU to find a path forward. For global companies, that creates an immediate product planning problem: an experience designed around Siri AI cannot yet reach European customers on those devices, and it is not clear when it will.

The strategic significance is distribution. Apple's competitors have pushed agents through chat apps and browsers. Apple is pushing them through the operating system, where they can reach into apps that people already trust with payments, health, travel and messaging. If developers adopt App Intents widely, Siri AI could become the layer that decides which app gets the customer's request.

For brands with significant iPhone audiences, the question is whether to be inside that layer or outside it. An app that exposes its key actions to Siri AI may capture requests that would otherwise go to a competitor or to the web. An app that does not may simply become invisible to a customer who no longer opens apps directly, and that is a decision worth making deliberately before spring.

AI Business Models Story 7 of 12

OpenAI Product Lead Says ChatGPT Has Reached 1.2 Billion Weekly Users and a Fifth of Google's Search Volume

ChatGPT's growth has not slowed. Adam Fry, the consumer product lead for ChatGPT at OpenAI, said at an OpenAI hosted dinner in New York that ChatGPT has reached 1.2 billion weekly active users, up from 1 billion in about 60 days, according to an account of the conversation published by Big Technology.

Fry also put a number on the comparison that matters most to the advertising industry. He said ChatGPT's search like query volume is roughly one fifth the volume of Google's queries. His definition is broad. "Every time you ask a question and go search the web, that's a search," he said, and the category includes lookups about people, places and things that need current information, as well as finance and health questions that call for web results.

The definition deserves scrutiny. A question that triggers a web lookup inside a chat assistant is not necessarily the same as a typed search on Google, and OpenAI's methodology is its own. Even so, the comparison signals where OpenAI sees its future: not as a novelty chatbot, but as a primary way people find information, the business at the heart of Google.

Fry also described a product change that will affect how hundreds of millions of people use the service. OpenAI plans to bring ChatGPT's Chat and Work modes together into one interface, removing the toggle users now choose between. The goal is a single assistant that infers whether someone wants a quick conversational answer or a longer agentic task. Asked when, Fry said only that it is coming soon.

That merger is harder than it sounds. Deciding automatically when to answer immediately and when to start a multistep job requires the model to judge intent reliably, and Fry acknowledged it requires further research. Getting it wrong in either direction is costly: a slow agent for a simple question frustrates users, while a shallow answer to a complex task wastes their time.

For marketers and digital leaders, the implications are immediate. If a meaningful share of information seeking now happens inside an assistant, visibility in AI generated answers becomes as important as ranking on a search results page. Brands should be testing how ChatGPT describes their products, their prices and their competitors, because for a growing share of customers, that answer is the first impression.

For Google, the number is a warning shot rather than a crisis. One fifth is far from parity. But at 1.2 billion weekly users, the alternative is no longer small.

Enterprise AI Story 8 of 12

Revolut Says Self Hosted Open Models Cut Its AI Coding Costs by Roughly 90 Percent

Revolut has given enterprise technology leaders a concrete case study in what happens when AI coding tools meet a finance department. David Abitbol, the fintech's ops and infrastructure lead, told the AI Engineer conference in New York that Revolut cut its AI coding inference costs by roughly 90 percent in under a year by making self hosted models its default, according to The Register.

The path was not immediate. Revolut rolled AI coding tools out to roughly 2,000 engineers, using commercial assistants first. Abitbol said spending had climbed with little measurable return. The company now serves GLM-5.3, an open weight model, on 56 Nvidia B200 GPUs.

The economics, as Revolut described them, are stark. Abitbol said self hosting cost about $0.30 per million input and output tokens, roughly a tenth of what it paid to use OpenAI's Codex or Anthropic's Claude Code.

Revolut did not abandon commercial models entirely. Because GLM is text only, the company routes requests that involve images to commercial models, while GLM handles code generation. That hybrid design is the real lesson: the default changed, not the entire stack.

This is a pattern worth watching across large engineering organizations. Commercial coding assistants are priced for convenience and frontier capability. For a company with thousands of engineers running agents all day, token costs can scale faster than headcount. Open weight models that are good enough for most routine coding work change that equation, provided the company has the infrastructure talent to run them.

There are real caveats. Running a GPU cluster carries capital costs, operational overhead and security responsibilities that a subscription does not. Regulated firms also need to weigh the provenance of the models they deploy, including where they were trained and how their weights have been vetted. And a model that is good enough today must keep pace as commercial rivals improve.

For chief technology officers, Revolut's numbers are a prompt to audit rather than a template to copy. Which share of your AI coding spend goes to tasks a cheaper model could handle? How much of your token bill comes from a small number of heavy users? Revolut's answer was a routing policy, not a ban, and that approach is available to any company willing to measure first.

AI Research Story 9 of 12

Anthropic Releases FlashProve, an Open Source Way to Prove Which Model Produced an Answer

Anthropic has released research and code aimed at a question that will matter more as AI is embedded in contracts, regulation and critical systems: how can anyone prove which model actually produced a given answer? The system, called FlashProve, is an open source prototype for what Anthropic calls provable inference.

The idea is simple to state and hard to build. A model provider commits to a specific set of model weights, then produces a cryptographic proof that those committed weights generated particular output tokens for a particular prompt. A verifier can check the proof without ever seeing the weights. FlashProve targets Llama 4 Maverick, a 400 billion parameter mixture of experts model, and the proofs hide the weights and intermediate values, though not the model's architecture.

Anthropic was candid about the cost. The company said FlashProve generates proofs at around 110 tokens per second on a single H100 GPU, which is about 70 to 95 times the GPU time of running the model itself on vLLM, a widely used open source serving library. Checking a proof is far cheaper: Anthropic said it takes 17 seconds on 24 threads. The prototype's context is limited to 1,024 tokens, and it covers greedy decoding of an exact integer version of the weights rather than ordinary sampled output.

The code is released under the Apache 2.0 license and includes a fast verifier, a slower verifier formally verified in the Lean proof assistant, and a technical report. Anthropic described the work as a research prototype and intermediate progress, and it announced no production deployment.

Why should executives care about an experimental cryptography project? Because trust in AI supply chains currently rests on promises. When a company buys access to a model, it trusts that the provider is serving the version it advertised, not a cheaper or altered one. When regulators require that a certified model be used for a high risk decision, they have no way to verify compliance after the fact. Provable inference would turn those promises into something checkable.

The overhead figures explain why this is not ready for everyday use. Paying roughly 70 to 95 times the compute to prove an answer makes sense only for rare, high stakes outputs. But cryptographic techniques have a history of becoming dramatically cheaper once a working design exists, and publishing the code invites others to improve it.

For risk, legal and procurement leaders, FlashProve is a signal of where assurance is heading. Contracts that today say a vendor will use a specific model may one day require proof that it did. It is worth understanding the concept now, before it appears in a request for proposal.

Funding & Investment Story 10 of 12

Traversal Raises $120 Million at a $1 Billion Valuation to Let AI Investigate Production Outages

Traversal, a startup building AI that investigates software outages, announced a $120 million Series B at a $1 billion valuation. Kleiner Perkins, Redpoint Ventures and Khosla Ventures jointly led the round, and total funding now stands at $173 million, according to Unite.AI's report on the announcement.

The company works on one of the most expensive problems in modern technology operations. When a production system fails, engineers on call must sift through logs, metrics, code changes and deployment histories to find the cause, often under intense pressure while customers are affected. Traversal's approach is to build a continuously updated model of a company's production environment that AI can read, then use a search engine it describes as causal to test thousands of hypotheses in parallel and arrive at evidence backed root causes.

The growth figures, as reported by Unite.AI, show why investors are paying attention. Traversal's monthly investigations grew from 513 to nearly 73,000 during 2026. Customers include Nubank, Harvey, DigitalOcean and PepsiCo, a mix that spans financial services, legal technology, cloud infrastructure and consumer goods.

The company is also connecting its system to the coding agents that are rapidly changing how software gets written. Through Traversal MCP, its production knowledge can be fed to coding agents, so that an agent proposing a fix can see how a system actually behaves in production rather than guessing from source code alone.

The funding fits a broader pattern. As AI writes more code, the volume of change flowing into production rises, and so does the potential for failures that no single engineer fully understands. Investors are betting that the operational side of software, including monitoring, debugging and incident response, will need its own AI layer to keep up with AI generated code.

There are hard questions for buyers. An AI that investigates incidents needs broad access to telemetry, code and internal documentation, which raises security and data governance concerns. Its conclusions also need to be trustworthy enough that engineers act on them at three in the morning. Tools in this category earn adoption by being right, consistently, when it matters.

For technology leaders, the practical move is to measure the problem before buying a solution. How long does your team take to find the root cause of a serious incident, and how much revenue does that time cost? If the answer is uncomfortable, the AI reliability category is now mature enough, and well funded enough, to deserve a serious evaluation.

Enterprise AI Story 11 of 12

Fiserv Doubles Its Savings Target to $1 Billion and Leans on AI Agents After a $6.45 Billion Write Down

Fiserv, the payments and banking technology provider, is betting that AI can help fix a cost structure built up over decades of growth. The company doubled its Project Elevate savings target to $1 billion as it reported third quarter 2026 results that included a $6.45 billion goodwill impairment, according to PYMNTS' account of the earnings call.

The restructuring is already well underway. Fiserv has reduced its workforce and contractor base by 3,000 since Takis Georgakopoulos became chief executive, and it has cut senior management by more than 10 percent, according to the same report. Georgakopoulos described an AI everywhere mindset across the company and said Fiserv has increasing confidence that it can drive greater efficiency.

The company offered a concrete example of what that looks like in operations. According to PYMNTS, Fiserv said AI agents working in the risk hold process for its Clover business cut decision times by two thirds and lowered false positives by more than 40 percent. Risk holds are the moments when a payment processor pauses a merchant's funds over suspected fraud or credit risk. Faster, more accurate decisions there mean fewer legitimate merchants left waiting for their money and fewer fraudulent transactions slipping through.

That example captures where AI is creating measurable value inside large financial companies. It is not in flashy customer chatbots, but in the high volume judgment calls that once required armies of analysts working through queues. Each decision is small, but at Fiserv's scale, shaving time and errors from millions of them adds up to real money and better customer experience.

The write down is the other half of the story. A goodwill impairment of that size signals that some past acquisitions are worth less than Fiserv paid for them, and it frames the cost program as a response to pressure, not only an opportunity. Investors will judge Project Elevate on whether savings show up in margins without damaging service quality for the banks and merchants that depend on Fiserv's systems.

For executives in financial services, Fiserv's playbook is instructive. Start with processes that are rules heavy, high volume and measurable, where an AI agent's performance can be compared directly against the human baseline. Report results in operational terms, such as decision time and false positive rates, rather than vague productivity claims. And recognize that headcount reductions will be part of the conversation, which requires a clear plan for the people affected and for the oversight of decisions that agents now make.

Industry Dynamics Story 12 of 12

U.S. Export Licenses Stall SK hynix's China Expansion and Cloud a $45 Billion New York Chip Plan

A plan to build major memory chip factories in upstate New York has run into the logic of U.S. export controls. According to people familiar with the matter cited by The Korea Times, SK hynix's Solidigm plans to invest about $45 billion to build two NAND flash memory plants in Marcy, New York. But the U.S. Commerce Department has not approved export licenses for the American chipmaking equipment SK hynix needs to expand its plant in Dalian, China.

The two are linked. SK hynix has argued that expanding production in Dalian would help fund its expansion in the United States. Without the licenses, that financing logic weakens, and the timeline for the New York project becomes harder to predict. Solidigm has said it is exploring options for building a NAND memory fab at potential locations around the world and has made no final decision.

NAND flash is the memory used for storage, and it has become a meaningful part of the AI infrastructure story. Training and serving large models requires enormous volumes of fast storage alongside the high bandwidth memory that sits next to AI accelerators. Demand from data centers has tightened supply across the memory market, which makes any delay in new capacity a concern for buyers.

The situation captures a tension at the heart of U.S. industrial policy. Washington wants foreign chipmakers to build advanced factories on American soil. At the same time, it is determined to limit the flow of advanced manufacturing technology into China. When a single company's global finances tie those goals together, policymakers are forced to choose which priority wins, and the company is left waiting.

For executives, the story is a reminder that semiconductor supply is increasingly shaped by licensing decisions as much as by engineering or demand. A fab announcement is not capacity until the approvals, equipment and financing all line up, and any one of those can stall in a government office.

For companies planning large AI deployments, the practical response is to treat memory and storage as strategic purchases, not commodities. That means longer term supply agreements, diversified vendors and realistic assumptions about when announced capacity will actually arrive. The Marcy project may well proceed. But until the licensing question is resolved, it remains a plan, and procurement teams should price it as one.