Computing News, March 1

Meta–AMD mega deal to diversify AI compute

Meta agreed to buy up to $60B of AMD AI accelerators and related CPUs over five years, a scale that implies “data-center power planning” as much as it does chip shopping. The structure is unusual: Meta also gets warrants that could translate into roughly a 10% stake if technical and shipment milestones are hit. AMD says the first wave starts in late 2026 and totals about 6GW of compute across generations, centered on next-gen Instinct parts (MI450-class) plus custom server CPUs. The broader story: hyperscalers are trying to reduce Nvidia single-supplier risk, lock in capacity early, and co-design hardware to fit their internal racks, networks, and power envelopes—turning chip roadmaps into strategic partnerships rather than spot purchases. Watch whether “equity-for-capacity” becomes a template in 2026 procurement. (Reuters)

Nvidia locks in Meta as a “millions of chips” customer

Nvidia announced a multiyear supply agreement with Meta covering current Blackwell GPUs and future Rubin-generation parts, plus its Grace and Vera CPUs. That last piece matters: Nvidia is pitching Arm-based CPUs not just as GPU companions but as general data-processing engines for databases, memory-heavy services, and the “AI agent” era where inference dominates. For Meta, the deal is another hedge—keep buying Nvidia’s best GPUs while also exploring in-house silicon and alternative suppliers. For Nvidia, it’s a loud signal that it intends to compete on the full server platform (CPU+GPU+interconnect+software), not only on accelerators, as data centers chase efficiency per watt and per dollar. Expect more “platform deals” that bundle chips, networking, and software support into one commitment. (Reuters)

Nvidia’s CEO tees up a renewed CPU war

In investor messaging around its expanding product line, Nvidia framed CPUs as “back in fashion” as AI shifts from training huge models to deploying fleets of inference and agent workloads that need fast general-purpose processing alongside accelerators. The company argues that platform-level integration (CPU, GPU, networking, and software) is becoming the differentiator, putting it on more direct collision courses with Intel and AMD. The timing also reflects data-center operators rebalancing: GPU clusters are still growing, but so are needs for high-core-count CPUs, memory bandwidth, and efficient orchestration—especially for systems that run many small, latency-sensitive AI tasks. Watch for 2026 server roadmaps to emphasize whole-rack designs, CPU-GPU memory coherence, and power-aware scheduling as competitive features, not just raw FLOPS. (Reuters)

Anthropic pushes Claude into enterprise “workflows,” not just chat

Anthropic unveiled new plug-ins and connectors aimed at making Claude an orchestration layer inside common business tools—think Drive/Gmail-style sources plus functions that touch spreadsheets and slide decks. The pitch is that AI should sit where work already happens, pull from governed internal data, and automate repetitive steps for teams like HR, banking, and engineering. Markets reacted because this model threatens several software categories at once: analytics, back-office automation, and even some security tooling if AI handles triage and remediation. The more subtle implication for computing: enterprises are standardizing on “AI runtime + connectors + policies” as a new platform layer, which will drive demand for secure tool-calling, audit logs, evaluation harnesses, and on-prem / VPC deployments to keep data boundaries intact. Expect a wave of “AI integration engineering” budgets in 2026. (Reuters)

CoCounsel hits 1 million users: legal AI goes mainstream

Thomson Reuters said its AI legal assistant CoCounsel reached one million users, sparking a sharp share rally and shifting the conversation from “AI hype” to adoption curves. Unlike general chatbots, CoCounsel is wrapped around proprietary legal content and workflow tooling (research, document review, drafting), which the company argues creates a defensible moat. The milestone matters beyond law firms: it’s a case study in where generative AI monetizes today—vertical products with curated data, clear ROI, and compliance constraints. For computing teams, the takeaway is that retrieval, citations, permissioning, and domain-specific evaluation are now “core platform features,” not add-ons, and organizations will increasingly buy AI as part of specialized software stacks rather than as a standalone model subscription. Watch whether similar “content + workflow” plays accelerate in finance, healthcare, and tax. (Reuters)

Microsoft Patch Tuesday: six exploited zero-days in one month

Microsoft’s February security release addressed roughly 54–59 vulnerabilities, including six zero-days reported as exploited in the wild. Coverage highlighted Windows and Office attack paths that can be “low interaction” (a click or opening a crafted file) and then chained into deeper compromise. For defenders, the headline isn’t just the count—it’s the operational reality that multiple exploit-ready issues can land at once across endpoint, browser, and productivity software. If you run Windows fleets, treat this as a priority patch window: identify exposed Office file-handling paths, check email gateway controls, and make sure exploit mitigation (ASR rules, macros policies, Protected View) is enforced while patches roll out. Also verify patch compliance on remote and BYOD endpoints, because attackers typically pivot to the least-managed machines after a high-profile Patch Tuesday. (Tenable®)

Apple patches an exploited dyld zero-day across iOS/macOS

Apple shipped updates (e.g., iOS/iPadOS 26.3 and macOS Tahoe 26.3) fixing CVE-2026-20700, a memory-corruption bug in dyld (the dynamic linker). Apple said it had reports the issue may have been used in an “extremely sophisticated attack” against targeted individuals, and credited Google’s Threat Analysis Group. This is a reminder that “closed ecosystem” doesn’t equal “no zero-days”: modern mobile exploits often chain several bugs, and dyld sits on a critical boundary between apps and system frameworks. For organizations with iPhones and Macs, the practical move is straightforward—accelerate OS update compliance, especially for high-risk users (executives, journalists, developers), and review MDM policies that delay major updates for compatibility testing. If you run mixed fleets, align Apple patch urgency with Windows Patch Tuesday cadence instead of treating mobile as “later.” ( Apple)

AWS outage story raises red flags about autonomous “AI ops”

Reuters, citing the Financial Times and AWS comments, described a December disruption in one AWS region where a cost-management feature was knocked out for about 13 hours after an internal AI coding tool (named Kiro in the report) deleted and recreated an environment. AWS emphasized the incident was limited to a single service, but the governance lesson is broader: when AI tools can change production configs, they need guardrails comparable to humans—least privilege, staged rollouts, and strong “undo” paths. As cloud providers embed AI into internal operations and customer consoles, expect more focus on change-control for automation, model evaluation for agentic tooling, and postmortems that treat AI systems as first-class actors in reliability engineering. A practical benchmark: any “AI agent” should produce the same audit trail your SRE team demands from people. (Reuters)

ByteDance designs its own inference chip, talks manufacturing with Samsung

Reuters reported ByteDance is developing an AI chip optimized for inference and is in discussions with Samsung for manufacturing, aiming for samples by late March and production later in 2026. The reported plan—hundreds of thousands of units at scale—signals how far vertical integration has spread beyond U.S. hyperscalers: major content and social platforms want to control costs, latency, and supply by owning silicon tailored to their recommendation and generative workloads. The computing angle is that inference, not training, is becoming the volume driver; chips built for efficient, high-throughput serving (and for specific model architectures) can deliver big savings at the data-center level. Watch for more custom accelerators paired with region-specific foundry partnerships, and for software stacks (compilers, kernels) to become the true differentiator as hardware options multiply. (Reuters)

Quantum processors get a “speedometer” for drifting qubits

Researchers at the Niels Bohr Institute (University of Copenhagen) reported a real-time method to track fast fluctuations in superconducting qubit relaxation rates—about 100× faster than prior characterization. They used an FPGA-based controller running Bayesian estimation to update a qubit’s health on millisecond timescales, revealing dynamics that were previously invisible because experiments averaged over too long. This matters because quantum systems are often limited by their worst qubits; being able to detect when a qubit turns “bad” in real time opens the door to dynamic calibration, smarter scheduling, and more resilient error-correction experiments. The work (published in Physical Review X) underscores that quantum progress increasingly depends on tight classical control loops and hardware/software co-design—treating calibration as a continuous, software-defined process rather than a lab ritual done “between runs.” (nbi.ku.dk)