AI Infrastructure Story 1 of 12

ASML Posts a Record Quarter and Warns It Cannot Build Enough Machines for 2027 and 2028

ASML, the company that makes the lithography machines every leading chipmaker depends on, reported record third quarter results on Wednesday and delivered a message that matters more than the numbers themselves: the AI buildout is now running ahead of the factory capacity that feeds it.

Total net sales reached 11.7 billion euros for the third quarter of 2026, with net income of 5.0 billion euros and a gross margin of 57.5 percent. That is a sharp step up from the second quarter, when ASML reported net sales of 9.3 billion euros and net income of 2.9 billion euros. The company guided fourth quarter sales of 14.2 billion to 15.2 billion euros at a gross margin of around 57 percent, and now expects full year 2026 sales of 44 billion to 45 billion euros.

The more consequential passage came from chief executive Christophe Fouquet in the results statement. He said customer demand is being driven primarily by their views on the size and duration of AI related investments, and that ASML now expects its supply as currently planned for both 2027 and 2028 will not fully meet the demand of its customers. The company is investigating ways to increase capacity for 2028 beyond the 30 percent increase it had previously communicated. Fouquet added that some customers are expressing interest in placing orders for 2029 and even 2030.

For executives, the signal is about lead times rather than earnings. When the supplier of the most advanced lithography tools tells the market that its planned output for the next two years is already short, it implies that leading edge chip capacity, and therefore the supply of the most capable AI accelerators, will stay tight well beyond the current product cycle. Customers asking to book equipment four years out is the behavior of buyers who believe the constraint is structural, not cyclical.

That has practical consequences far downstream of the fab. Enterprises negotiating multiyear cloud and GPU commitments should assume that scarce capacity will continue to be allocated to the largest buyers first, and that price relief from a supply glut is not on the near horizon. Infrastructure plans that assume abundant, cheap frontier compute in 2027 deserve a second look.

It also reframes the competitive map. Chipmakers that secured ASML slots early gain an advantage that money alone cannot quickly replicate, because the machines themselves take years to build and install. Governments pursuing domestic semiconductor ambitions face the same queue.

The question for the quarters ahead is whether ASML can find the extra 2028 capacity it is now hunting for, and whether the AI investment cycle holds long enough to justify it. For now, the company at the very start of the AI supply chain is saying the demand is real, durable, and bigger than its current plans.

ASMLSemiconductorsEUVAI Demand

Policy & Regulation Story 2 of 12

More Than 400 Scientists, Led by Hinton and Bengio, Call for Air Gapped AI and a Global Containment Body

A group of more than 400 scientists, including Geoffrey Hinton and Yoshua Bengio, released an open letter on AI and cybersecurity on Tuesday, October 13, calling for frontier AI systems capable of serious cyber operations to be kept physically cut off from the internet by default while they are developed and tested.

The letter asks organizations to default to air gapping cyber capable models, meaning disconnecting them from outside networks, and to adopt high assurance containment systems immediately. It also calls for an international body to standardize containment practices for the public good and to track containment failures as they occur. Signatories also include Stuart Russell and Ron Rivest. BetaKit reported that nearly 70 of the signatories are from Canada.

What sets this letter apart from earlier warnings about AI risk is its focus. Rather than speaking in general terms about catastrophic outcomes, it zeroes in on a concrete engineering question: where a powerful model physically lives while people are still learning what it can do. That framing reflects a shift in the safety debate from abstract alignment concerns toward operational security, the same discipline that governs how banks protect their core systems or how governments handle classified networks.

The letter's signatories are effectively arguing that the industry should not wait for a serious containment failure before treating cyber capable models the way other high consequence technologies are treated.

For business leaders, the practical implication is that containment is likely to become a procurement and compliance question, not just a research topic. Enterprises that fine tune or host advanced models internally, particularly in security, defense, finance, and critical infrastructure, should expect regulators and insurers to begin asking how those systems are isolated, who can reach them, and how incidents are recorded.

An international body of the kind the letter describes does not yet exist, and building one would require governments to agree on standards at a moment when AI policy is sharply contested. But open letters signed by figures of this stature have historically shaped the agenda that regulators work from. The call for air gapping by default gives policymakers a specific, testable requirement to debate, which is more than most prior appeals have offered.

Whether labs adopt the practice voluntarily, or wait to be told, will be one of the clearer tests of how seriously the industry takes its own warnings about cyber capability.

AI SafetyCybersecurityOpen LetterGlobal Governance

AI Safety Story 3 of 12

MI5 Chief Says Terrorists Are Already Trying to Use AI, and Names Bioweapons as His Biggest Worry

The head of Britain's domestic security service used his annual threat update on Tuesday to put artificial intelligence near the center of the national security picture, telling an audience at Thames House that MI5 has caught terrorists seeking to exploit AI to improve their reconnaissance and attack methods.

Ken McCallum, the Director General of MI5, said that among the ways AI could be misused, the case to which he would attach the greatest weight is its potential to change the stakes on existing or even new bioweapons. He also announced that MI5 is setting up an AI Threat Assessment Centre, working with the National Crime Agency and the National Cyber Security Centre, to assess the risks posed by the technology as it develops.

The speech covered far more than AI. McCallum said MI5's caseload on state threats has quadrupled over the past ten years, reflecting a broader rise in hostile activity directed at the United Kingdom. But the decision to stand up a dedicated AI assessment unit signals that the security service now regards AI as a standing category of threat rather than a speculative one.

For executives, the remarks carry two messages. The first concerns capability uplift. A recurring concern among security officials is that AI may give relatively unsophisticated actors access to knowledge and planning help that previously required expertise. That risk sits squarely with the companies that build and deploy powerful models, and it is one reason governments are pressing for stronger safeguards, testing, and incident reporting from AI developers.

The second message is about the defensive side of the ledger. When a national intelligence agency publicly commits resources to understanding AI enabled threats, it typically foreshadows new guidance, information sharing arrangements, and expectations for the private sector. Companies in biotech, chemicals, critical infrastructure, and AI development should anticipate closer engagement from government on how their systems and data could be misused, and how quickly they report suspicious activity.

The new center also reflects a pattern visible across allied governments: AI risk is migrating from technology ministries into the security establishment. That shift tends to change the tone of regulation, emphasizing controls, audits, and accountability over voluntary commitments.

By naming bioweapons as the gravest misuse case and confirming that terrorists are actively experimenting with AI, he gave policymakers and boards a clear reason to treat AI safety as part of national resilience planning rather than a niche technical concern.

MI5National SecurityBioweaponsUnited Kingdom

Enterprise AI Story 4 of 12

Anthropic Replaced Git for Its Own Code Because Claude Was Writing Too Much of It

Anthropic disclosed on Tuesday that it built its own version control system, called Scion, and fully migrated its main code repository off git in five months. The reason is one of the clearest illustrations yet of how AI coding agents are reshaping the economics of software engineering inside the companies that build them.

In an engineering post, Anthropic said it averaged around 2,500 code changes merged per day in May 2026. By September, that figure had climbed to more than 30,000 per day. The company said the surge was driven by Claude's ability to write code, and that the volume overwhelmed infrastructure designed around human paced development. The company performs on the order of 25 million code checkouts per day for continuous integration alone.

Scion is the result. Anthropic said the new system's infrastructure footprint is 200 times smaller than the git mirroring infrastructure it replaced, and that a team of seven people built it and completed the migration.

The headline for technology leaders is not the tool itself but the scale shift behind it. A more than tenfold increase in merged changes over roughly four months is not a productivity gain that slots neatly into existing workflows. It breaks them. Systems that were sized for a world where every change was written and reviewed by a person, including source control, build pipelines, test infrastructure, and code review, become bottlenecks when agents generate work around the clock.

Most enterprises are earlier on this curve than Anthropic, which is unusually aggressive in using its own models. But the direction is the same. Companies rolling out coding agents at scale should expect pressure to show up first in their developer infrastructure: build queues lengthening, repository performance degrading, review backlogs growing, and cloud costs for continuous integration rising sharply.

That makes infrastructure readiness a strategic question, not a housekeeping one. Before measuring the return on AI coding tools, engineering leaders should model what happens to their pipelines if output rises by an order of magnitude, and decide where human review remains essential when the volume of machine written changes outpaces any team's ability to read them.

Anthropic's experience also hints at a new category of opportunity. If AI driven development routinely strains standard tools, vendors that rebuild version control, testing, and review for agent scale workloads may find a ready market among enterprises following the same path.

AnthropicDeveloper ToolsVersion ControlAI Coding

AI Business Models Story 5 of 12

Character.ai Files for Chapter 11, and Its Filings Expose the Hard Math of Consumer AI

Character.ai, once one of the most prominent consumer AI startups, has filed for voluntary Chapter 11 bankruptcy protection to pursue a sale of the business. CBS News reported that Disney has agreed to pay 20 million dollars for a non exclusive, perpetual license to the company's intellectual property.

Chief executive Colin Adams attributed the company's difficulties to significant legal, regulatory and operational headwinds. The company has said its service continues to operate while the process unfolds, and it expects the business to be sold by late November or early December.

The more revealing story is in the numbers. According to bankruptcy documents reviewed by 404 Media, Character.ai spent 56.1 million dollars on compute from January through August of 2026, while its total monthly revenue peaked at 5.3 million dollars. As of August, the documents show the company was losing roughly 8 million to 9 million dollars every month.

Those figures put a hard edge on a question that has hung over consumer AI for years. Chat products can attract enormous engagement, but every conversation carries a real inference cost, and heavy users are often the most expensive to serve. When monthly revenue is a fraction of monthly compute spend, growth does not fix the problem. It deepens it.

The legal pressures compounded the economics. The company's leadership has pointed to legal and regulatory headwinds as a central factor in its decline. Litigation and compliance costs are now a material line item for companion and entertainment AI products, alongside compute.

For executives evaluating consumer facing AI, there are three lessons. First, unit economics must be modeled per conversation, not per user, because usage intensity varies enormously. Second, monetization has to be designed in from the start; advertising and subscriptions layered onto a free, compute heavy product struggle to close the gap. Third, trust and safety obligations, especially where minors may be present, are not optional overhead but a determinant of whether a product survives.

Disney's license deal suggests that the value left in Character.ai may lie less in the service itself than in its technology and character design capability. Whoever buys the business will inherit both a loyal user base and the cost structure that brought it to court.

Character.aiBankruptcyConsumer AIDisney

Industry Dynamics Story 6 of 12

Huawei Closes In on Nvidia in China's AI Chip Market, IDC Data Shows

Nvidia's grip on China's market for AI chips weakened sharply in the first half of 2026, while Huawei moved close to parity, according to an IDC report seen by Reuters.

IDC's figures show Nvidia shipped 678,000 AI accelerators in China between January and June, down 44 percent from a year earlier, giving it a 34 percent share of the country's data center AI chip market. Huawei shipped 572,000 accelerators over the same period, for a 28 percent share. Total AI accelerator shipments in China reached 2.008 million units in the half, according to the report.

The shift reflects two forces working in the same direction: American export restrictions on advanced chips, and a push inside China toward domestic suppliers.

For Nvidia, a 44 percent drop in Chinese shipments is significant even against strong demand elsewhere. Every unit ceded there becomes a unit that helps a domestic rival refine its hardware and software stack. Huawei's progress matters not just for the volume it represents, but for the ecosystem of developers who learn to build on its platform instead.

The broader implication for executives is that the global AI hardware market is splitting into two increasingly separate spheres. Multinationals that operate AI workloads in China should plan for a supply environment dominated by domestic accelerators, with different software tooling, performance characteristics, and support models. Model portability across hardware platforms becomes a practical requirement rather than a nice to have.

It also matters for investors and partners outside China. A domestic Chinese chip sector that ships hundreds of thousands of accelerators per half year is building manufacturing experience and scale that could, over time, compete internationally in markets not bound by American restrictions.

Nvidia still leads in China by IDC's count. But the numbers show that its lead is no longer a moat. It is a margin that is shrinking quickly, and policy on both sides of the Pacific is accelerating the trend.

NvidiaHuaweiChinaAI Chips

AI Infrastructure Story 7 of 12

Intel Names Its Inference Chip the GPU D Series and Bets on Memory Over Raw Power

Intel gave its data center inference accelerator, previously codenamed Crescent Island, a product name this week: the Intel GPU D Series. The announcement, made around the Open Compute Project Global Summit, is Intel's latest attempt to win a place in an AI hardware market that has been dominated by Nvidia.

Intel's product page lists memory of up to 480GB of LPDDR5X per card in a 350 watt, air cooled design, built on 32 Xe3P cores and 256 XMX matrix engines. The chip delivers 1.3 petaflops of FP8 compute and 1.5 terabytes per second of memory bandwidth, and Intel said customer sampling is underway. StorageReview reported that the 480GB figure applies to partner built cards, while Intel's own branded card carries 160GB.

The design choices tell the strategy. Rather than chasing the highest possible training performance, Intel is aiming squarely at inference, the work of running models once they are trained, which is where enterprise AI spending increasingly concentrates. Large memory capacity matters for inference because it lets a single card hold bigger models and longer context windows, reducing the need to split work across many expensive accelerators.

Using LPDDR5X, a memory type more common in laptops and phones than in flagship AI chips, is the notable trade off. It is cheaper and more power efficient than the high bandwidth memory used in leading accelerators, but it moves data more slowly. Intel is betting that for many inference workloads, especially agent systems and retrieval tasks that keep large amounts of information in memory, capacity and cost per card matter more than peak bandwidth.

The 350 watt, air cooled envelope is equally deliberate. Many enterprise data centers cannot easily accommodate the liquid cooling and extreme power draw of the newest flagship accelerators. A card that fits into conventional servers lowers the barrier for companies that want to run AI on their own infrastructure rather than renting it.

For technology buyers, the D Series adds a credible option to a market where supply and pricing have been dictated by a single vendor. Its success will depend less on the specifications than on software maturity, partner availability, and real world performance on the models enterprises actually run.

The D Series is a narrower, more pragmatic bet than a head to head fight for training workloads. If inference becomes the dominant AI workload, a chip built for memory heavy, cost sensitive deployments could find room in a market that has so far left little space for challengers.

IntelInferenceGPUsOCP Summit

AI Safety Story 8 of 12

Google DeepMind and Singapore Put a Scam Detector Directly on the Phone

Google DeepMind and GovTech Singapore, the city state's government technology agency, have built an AI system that detects scam conversations directly on a user's phone, a design meant to fight fraud without sending private messages to the cloud.

The system, presented at Singapore International Cyber Week, runs in two stages. A lightweight classifier first screens incoming messages using minimal power. When something looks suspicious, a lightweight Gemma model fine tuned on local scam conversations examines a larger portion of the exchange to confirm the threat or dismiss a false alarm. If manipulation is detected, the user receives a warning on the device urging them to pause before sharing money or sensitive information.

The scale of the problem explains the effort. According to Google DeepMind, scams make up over half of Singapore's reported crimes, and losses have exceeded 4 billion Singapore dollars since 2019. Complex social engineering scams, in which criminals build trust over time before striking, account for up to 80 percent of those losses.

Mark Chen, Director of Trust and Safety at GovTech Singapore, summed up the design principle: "Fighting scams shouldn't mean sacrificing personal privacy." Running the model on the device matters because many messaging platforms use end to end encryption, which prevents any server from inspecting conversations. The only place a scam can be detected in real time, without breaking that encryption, is on the phone itself.

For executives, the project is a template worth studying. It shows how small, specialized models can deliver protection that large cloud models cannot, precisely because they run where the data already lives. The same architecture applies to banks, telecom operators, and payment providers that want to flag fraud in real time while respecting privacy rules and encryption.

It also illustrates a growing model of public and private collaboration. Governments hold the local data on how scams actually unfold, while AI labs hold the modeling expertise. Combining the two produced a tool tuned to the specific tactics criminals use in Singapore, something a generic global model would struggle to match.

The two stage design deserves attention too. By reserving the more capable model for messages that already look suspicious, the system keeps battery and processing costs low enough to run continuously on ordinary smartphones. As fraud increasingly relies on AI generated messages and voices, defenses that live on the device and respond in the moment may become a standard expectation for consumers and regulators alike.

Google DeepMindSingaporeFraudOn Device AI

AI Safety Story 9 of 12

Anthropic and OpenAI Sought Control Over What Outside AI Evaluators Can Publish

Anthropic and OpenAI pushed for substantial control over what outside evaluators of their AI models are allowed to publish, citing concerns about trade secrets, Reuters reported on Tuesday. The report lands at a sensitive moment, as both companies have publicly promoted independent testing as a cornerstone of responsible AI development.

The tension centers on a new kind of arrangement in which outside experts work closely with labs to probe their models before and after release. In September, Faculty, the AI unit of Accenture, was named as Anthropic's first embedded external evaluator. Faculty's chief executive, Marc Warner, told Reuters that evaluators are only useful if people can trust what they report, and that independence and transparency are critical.

The core dilemma is real. Evaluators who work inside a lab see unreleased models, internal tooling, and training details that are commercially valuable. Labs have a legitimate interest in keeping that material confidential. But if the lab being evaluated can shape or veto what the evaluator says publicly, the evaluation loses much of its value as an independent check.

For executives, this matters because third party evaluations are increasingly what boards, regulators, and customers rely on to judge whether an AI system is safe to deploy. Enterprise buyers in regulated industries often point to external testing when justifying the adoption of a frontier model. If those reports are subject to editorial control by the vendor, buyers need to know where the line sits between redacting secrets and softening findings.

The episode also highlights a gap in governance. There are not yet widely accepted rules for how independent AI evaluations should work, who pays for them, what evaluators can disclose, and how disputes are resolved. In that vacuum, terms are negotiated privately between labs and individual testing firms, with the public seeing only the final report.

That is likely to draw regulatory interest. Governments building AI oversight frameworks have leaned heavily on the idea of external evaluation, and evidence that labs sought publication control will strengthen the case for formal standards, possibly including requirements that evaluators disclose when material was removed from a report.

For companies procuring frontier AI, the practical step is straightforward: ask vendors how their external evaluations are governed, whether evaluators can publish findings without approval, and how redactions are disclosed. The answers will say a good deal about how much weight those evaluations deserve.

AnthropicOpenAIAI EvaluationTransparency

Funding & Investment Story 10 of 12

Causa Prima Raises $20 Million to Let AI Agents Settle Business Payment Disputes

Causa Prima, a startup building AI agents that negotiate with each other on behalf of businesses, has raised a 20 million dollar seed round led by Base10 Partners, bringing its total funding to 30 million dollars, Fortune reported.

The company's software lets AI agents representing buyers and suppliers work together to resolve billing disputes and agree on payment terms, with the aim of speeding up payments between companies. Fortune reported that Causa Prima's leadership includes Maex Ament, who founded Taulia, Henrik Gebbing, a cofounder of the fintech Finoa, and chief technology officer Philip Stanislaus. The company has 25 employees across Madrid and Munich, according to the report.

The pitch targets one of the least glamorous but most expensive corners of corporate finance. Business to business payments routinely stall over mismatched invoices, disputed line items, and slow approvals, tying up working capital for both sides. Resolving those issues today usually means emails, phone calls, and manual reconciliation by accounts payable and receivable teams.

Causa Prima's bet is that much of that back and forth can be handled by agents acting within rules set by each company. According to Fortune, the startup makes money by taking a cut when its agents successfully negotiate early payment between a buyer and a supplier, aligning its revenue with measurable cash flow improvements for customers.

The round is notable for what it represents beyond a single company. Most enterprise AI agents today work inside one organization, automating internal tasks. Causa Prima is building for a world where agents from different companies interact directly, each representing its own side's interests. That raises new questions about trust, authorization, and auditability that finance leaders will need to answer before letting software commit their company to terms.

For chief financial officers, the opportunity is tangible: faster dispute resolution and earlier payment can free working capital without new borrowing. The risks are equally concrete. Agents negotiating on a company's behalf need clear limits on what they can concede, and every agreement they reach must leave a record that auditors and counterparties can verify.

The founders' backgrounds in invoice financing and fintech suggest a team that understands those constraints. Whether agent to agent negotiation becomes a standard part of the order to cash cycle will depend on how quickly enterprises grow comfortable letting software speak for them in money matters.

Agentic AIPaymentsSeed FundingB2B Finance

AI Models Story 11 of 12

Liquid AI Releases a 24 Billion Parameter Open Weight Model Built to Run Agents on a Laptop

Liquid AI released LFM2.5-24B-A2B on Tuesday, an open weight model with 24 billion total parameters but only about 2 billion active for any given piece of text, designed to run fast, private AI agents on ordinary computers instead of in the cloud.

The model uses a mixture of experts design, in which only a small slice of the network does work on each token. That keeps computation low while preserving the knowledge stored in a much larger model. It supports a context window of 131,072 tokens. Liquid AI scaled pre training from 17 trillion to 33 trillion tokens compared with its earlier generation.

Speed is the headline. Liquid AI says the model runs at 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, rates fast enough for interactive agents that call tools, read documents, and take multiple steps without a cloud connection.

On quality, Liquid AI reports that LFM2.5-24B-A2B scores 48.09 on the LiveClaw agent benchmark, compared with 38.19 for Qwen3.6-35B-A3B, a larger model. The company's own figures also show where the model trails: Qwen3.6-35B-A3B scores higher on SWE-Bench Verified, a software engineering test, and on GPQA Diamond, a test of graduate level science questions. Those are vendor reported results, and independent testing will matter.

For enterprises, the significance lies in where AI can run. Many organizations want agents that can work with sensitive documents, customer data, or proprietary code without sending that material to a third party's servers. A capable model that runs locally on high end laptops and workstations makes that possible, while also eliminating per token cloud charges for routine tasks.

The release fits a broader trend toward efficient, specialized models that sit alongside, rather than replace, the largest frontier systems. A sensible architecture for many companies will route simple and sensitive work to local models and reserve expensive cloud models for the hardest problems.

Liquid AI's results also underline how quickly the efficiency frontier is moving. A model that activates about 2 billion parameters per token and competes with larger peers on agent tasks suggests that the hardware already sitting on employees' desks is becoming a viable place to run meaningful AI work. Technology leaders planning their AI infrastructure should factor that shift into decisions about devices, licensing, and data governance.

Liquid AIOpen Weight ModelsOn Device AIAI Agents

AI Infrastructure Story 12 of 12

Nvidia Opens Its Vera Rubin Rack Designs and Joins Rivals on a Common Data Center Blueprint

Nvidia used the Open Compute Project Global Summit this week to push more of its AI factory engineering into the open, contributing designs for its next generation Vera Rubin systems and signing on to a shared data center blueprint with some of its biggest competitors and customers.

Nvidia said it contributed Vera Rubin rack and compute tray designs through its open MGX architecture, and that the MGX 1.2 specification supports racks of up to 250 kilowatts with a liquid cooled 5,000 amp busbar. The company also said it had joined AMD, Google and Meta in publishing version 0.7 of the Open Data Center Specification, which covers power architecture, cooling and physical layout for AI facilities.

Nvidia and Cisco announced plans for USFP, a new optical transceiver form factor with a path to 400 gigabits per second per lane, and Nvidia said Meta is adopting its Spectrum-6 Ethernet switches, which offer 102.4 terabits per second of switching capacity.

The numbers capture how radically AI has changed the physical data center. A rack drawing 250 kilowatts consumes many times more power than a conventional enterprise server rack, and moving that much electricity and heat requires heavy busbars, liquid cooling, and buildings designed from the ground up for the load. Standardizing those designs helps the entire supply chain, from power equipment makers to construction firms, build faster and with fewer custom engineering decisions.

The joint specification with AMD, Google and Meta is the more strategically interesting move. Nvidia and AMD are direct competitors in AI accelerators, and Google and Meta are building their own chips. Agreeing on a common facility blueprint suggests the industry sees the building itself, its power and cooling, as shared infrastructure, even as competition over the chips inside intensifies.

For executives planning AI capacity, the message is that facility design is becoming as important as chip selection. Organizations building or leasing data center space for AI should ensure that sites can support liquid cooling and very high power densities, and that designs align with emerging open standards to preserve flexibility across hardware vendors.

Open hardware specifications have historically lowered costs by widening the pool of suppliers. If the same dynamic plays out for AI facilities, the beneficiaries will include cloud providers, colocation operators, and large enterprises that want to run AI on infrastructure they control. Nvidia, for its part, gains by making its own platform the reference point that everyone else builds around.

NvidiaOCP SummitData CentersOpen Hardware