Introduction
This was not a quiet week for AI — it just looked quiet on the surface. Underneath, something more interesting was happening: the technology and the world it runs on are starting to pull in opposite directions.
On the software side, the pace is striking. Anthropic acquired Vercept to push Claude’s ability to see and operate software interfaces past the 72.5% mark on OSWorld — up from 15% just fifteen months ago. Cloudflare reimplemented 94% of a major web framework in a single week using Claude, for roughly the price of a cheap flight. Google launched Nano Banana 2 (Gemini 3.1 Flash Image) and Perplexity had it running inside their new multi-agent Computer tool on the same day — a day-zero integration that would have been unimaginable two years ago. The software layer is moving fast and integrating even faster.
On the physical side, the signals tell a different story. Eight hyperscalers are on track to spend $710 billion on AI infrastructure in 2026 — and that capital race is already raising the price of RAM in consumer laptops and potentially shrinking the memory in your next budget smartphone. Power grids cannot keep up with the demand; hyperscalers are increasingly bypassing them and funding nuclear reactors instead. And this week, people in the UK took to the streets to protest the expansion of data centres that strain local power grids.
Somewhere between a 3D-stacked Fujitsu CPU shipping its first samples and protesters marching outside OpenAI’s London offices, the shape of this week comes into focus. Here is what stood out to me most.
1. Fujitsu’s Monaka CPU Ships First Samples Using Broadcom’s 3D Chip Tech
Fujitsu taps Broadcom's 3D chip tech for 144-core Monaka CPU
The Register reports that Fujitsu’s upcoming 144-core Monaka CPU will be built using Broadcom’s XDSiP (Extreme Dimension System in Package) 3D chip-stacking technology, and that the first samples have already shipped this week. The Monaka design stacks four 2nm compute dies — each with 36 Armv9 cores — alongside SRAM chiplets on a 5nm process, all interconnected via a central I/O die with 12 channels of DDR5 and PCIe 6.0. Broadcom’s VP of ASIC products confirmed that Monaka is one of roughly half a dozen designs in development using this platform.
Why This Matters
This is the kind of story that is easy to overlook when the headlines are full of new model releases, but it deserves attention. Advanced packaging — stacking different chips together at very high density — is quietly becoming as strategically important as the compute itself. Fujitsu openly disclosing its collaboration with Broadcom on this is unusual; most of these chip customers are notoriously tight-lipped. When a partner starts talking publicly, it usually means the technology is mature enough to be competitive. The Monaka chip is currently targeted for launch around 2027, so we are watching a longer-horizon bet here, but the direction is clear: custom, tightly integrated silicon designed for specific workloads.
2. AWS Upgrades Its Large Model Inference Container Stack
AWS added new capabilities to its LMI (Large Model Inference) container, including LMCache integration for disaggregated prefilling and KV-cache offloading, as well as support for newer NVIDIA hardware and TensorRT-LLM paths. AWS also reports benchmark gains, including improvements to time-to-first-token and throughput.
A quick note on terms: disaggregated prefilling separates the compute-heavy “prefill” phase — where the model processes your input prompt — from the “decode” phase, where it generates tokens. Splitting these apart lets you allocate hardware more efficiently rather than leaving GPUs idle during the cheaper part of the job.
Why This Matters
The optimisation race is shifting to inference plumbing. Teams that improve latency and utilisation at the container and runtime level can deliver meaningful gains without waiting for the next frontier model release. It is a bit like tuning the engine rather than buying a new car — less glamorous, but often exactly what production systems need.
3. TrendForce Expects Another Big Jump in Hyperscaler Spend
Top cloud providers to outspend Ireland's GDP on AI in 2026
According to The Register’s coverage of TrendForce estimates, combined capex by the eight largest cloud providers — Google, Amazon, Meta, Microsoft, Oracle, Tencent, Alibaba, and Baidu — is projected to exceed $710 billion in 2026, up roughly 61% year-over-year. The four largest US players (Google, Amazon, Meta, Microsoft) alone account for approximately $635 billion of that total. On the AI server side, the spend surge is driving a notable memory shortage, as chipmakers shift manufacturing capacity toward high-margin high-bandwidth memory (HBM) used in GPU servers.
Google remains the only hyperscaler deploying more ASIC-based servers than GPU-based ones, with its Tensor Processing Units (TPUs) expected to feature in around 78% of AI servers shipped to Google datacenters this year.
Why This Matters
This is a scale signal, not a hype signal. $710 billion is, as The Register cheerfully points out, more than Ireland’s entire GDP. The capital plan suggests hyperscalers still expect sustained AI demand and are budgeting accordingly, even as power, procurement, and deployment constraints remain tight.
The memory shortage rippling out from this deserves a closer look, because it reaches well beyond server racks into your everyday devices. As manufacturers reallocate advanced production capacity toward HBM and server DRAM, TrendForce dramatically revised its Q1 2026 price forecast upward — from an initial 55–60% QoQ increase estimate to 90–95%, calling it an unprecedented single-quarter adjustment. PC DRAM contract prices are now projected to more than double QoQ, setting a new historical record. For consumer hardware, DRAM and NAND Flash are expected to exceed 20% of a notebook’s total bill-of-materials in 2026. Some Android brands are already downgrading base model specs and raising launch prices; budget notebooks risk delayed replacement cycles as consumers push back. The base model smartphone — which once came with 6–8 GB RAM as standard — may return to 4 GB in 2026 as brands cut costs to survive.
The takeaway: Big Tech’s AI spending is effectively a hidden tax on your next laptop or phone. HBM demand is cannibalising the silicon needed for consumer devices — PC DRAM prices are on track to more than double in a single quarter, a historical first. The servers and the budget PC market are competing for the same silicon, and right now, the servers are winning.
4. Microsoft Investigates High-Temperature Superconductors for Data Centre Power
High-Temperature Superconductors Could Redefine Data Center Power Density — TechRepublic
Microsoft touts immature HTS tech for datacenter efficiency — The Register
TechRepublic reports that Microsoft is exploring high-temperature superconductor (HTS) cables as a potential solution to the data centre power bottleneck. Unlike conventional copper and aluminium wiring, HTS cables conduct electricity with zero resistance, meaning no energy is lost as heat and no voltage drop over distance. Microsoft has already demonstrated a prototype server rack powered directly by a 3MW superconducting cable — built with VEIR, a Microsoft-backed startup working on HTS power delivery systems. The cables can carry five times more current over twenty times less space than copper equivalents, and they operate at around -196°C using liquid nitrogen cooling, which is significantly more accessible than older superconductors that required temperatures close to absolute zero.
The honest caveat: a Microsoft spokesperson told The Register that “HTS remains in the development and evaluation stage for adoption at Microsoft’s scale,” and that the current focus is on “testing, validating and building confidence in the technology with partners.” VEIR has said it is moving toward commercial deployment in 2026, but full data centre rollout remains some years away.
Why This Matters
The power problem is becoming the defining constraint of the AI era — and this story illustrates why the solutions are harder than they look. HTS cables are real, they work, and Microsoft has prototype hardware running. But the path from a factory test to widespread deployment runs through materials availability, cooling infrastructure, cost reductions, and utility standards that do not yet exist at the required scale. It is worth keeping both facts in view: the technology is genuinely promising, and it is genuinely far from ready. In the meantime, the grid constraints it is meant to solve are not waiting.
5. Canada’s Privacy Regulator Presses OpenAI on ChatGPT Data Handling
Canadian government demands safety changes from OpenAI
Engadget reports that Canada’s Privacy Commissioner said OpenAI’s proposed ChatGPT changes satisfied concerns related to the collection, use, and disclosure of personal information.
Why This Matters
Regulatory pressure is increasingly translating into concrete product and policy changes. That matters for developers and teams that rely on external AI APIs and need a predictable compliance posture across jurisdictions. The fact that proposed changes were enough to satisfy a regulator is also a small but real sign that some regulators and AI companies are finding ways to work together, rather than simply talking past each other.
It is worth noting, though, that Canada’s review focused specifically on data handling — collection, use, and disclosure. It did not address the separate and serious question of psychological safety. That concern is moving through the courts: a wrongful death lawsuit filed in December 2025 alleges that ChatGPT’s sycophantic responses reinforced a user’s paranoid delusions in the lead-up to a killing. Privacy compliance and safe behaviour are not the same thing — and regulators have not yet caught up to the gap between them.
Apps & Tool Updates
1. Microsoft Copilot to Auto-Launch in Edge When Opening Outlook Links
Microsoft to auto-launch Copilot in Edge whenever you click a link from Outlook
The Register reports that Microsoft has announced a new behaviour for Edge: whenever you open a link from Outlook, the Copilot side pane will automatically open alongside it. The feature appeared on the Microsoft 365 roadmap on February 25, with a rollout expected to begin in May 2026. According to Microsoft, it is designed to provide contextual insights based on email and destination content.
Whether it will be opt-in or opt-out has not yet been confirmed. The Vivaldi browser CEO, quoted in the article, was not exactly thrilled about it — raising fair concerns about corporate email privacy and the wisdom of having an LLM automatically reading your messages as you browse.
Why This Matters
Assistants are moving from an optional feature to a default UI layer. This increases adoption potential, but it also raises new expectations for user control, transparency, and interruption management — especially for enterprise administrators already playing an ongoing game of Whac-A-Mole to manage Copilot’s reach across Microsoft’s product suite. If you manage Office 365 policies for a team, this one is worth tracking.
2. Cloudflare Reimplements Most of the Next.js API With Claude in One Week
Cloudflare experiment ports most of Next.js API 'in one week' with AI
The Register reports that Cloudflare engineering director Steve Faulkner used Anthropic’s Claude — via Claude Code and the OpenCode agent — to reimplement 94% of the Next.js API in roughly one week, spending approximately $1,100 on API tokens across more than 800 development sessions. The motivation was not to show off AI coding for its own sake, but to address a genuine lock-in problem: Next.js tooling is deeply tied to Vercel’s infrastructure, making it difficult to deploy to other platforms such as Cloudflare, Netlify, or AWS Lambda without significant reshaping. The result — an open-source project called Vinext — replaces Next.js’s Turbopack build chain with Vite 8 (powered by the new Rust-based Rolldown bundler) and produces client bundles that are around 57% smaller, with build times up to 4.4 times faster. The project ships with over 1,700 Vitest unit tests and 380 Playwright end-to-end tests ported directly from the Next.js test suite, and several production sites, including CIO.gov, are already running on it.
Faulkner was clear that the human role remained critical throughout: he spent several hours upfront defining the architecture with Claude, then directed the implementation piece by piece and course-corrected along the way. Almost every line was written by the AI — but none of it was unsupervised.
Why This Matters
Think of this as the most concrete proof-of-concept yet for AI-assisted legacy code modernisation as a service. A framework reimplementation that “would normally take a team of engineers months, if not years” (Faulkner’s words) was completed by one person with a well-crafted prompt and about the price of a cheap flight. The conditions that made it possible — a well-documented target API, a comprehensive existing test suite, and a model capable of holding the full system in context — are conditions that describe a huge category of enterprise software: aging internal tools, framework migrations, and vendor lock-in problems that teams have been deferring for years. If you have a legacy migration sitting in your backlog, this week’s news is a reason to revisit the estimate.
3. Samsung’s Galaxy S26 Adds AI Call and Privacy Controls
Samsung unveils Galaxy S26 lineup with AI-heavy software updates
The Register highlights Samsung’s Galaxy S26 launch, featuring AI-focused additions, including AI call handling options and privacy-focused display controls to reduce the risk of shoulder surfing.
Why This Matters
Consumer AI differentiation is moving toward trust and control features, not only model quality. That shift is likely to influence enterprise expectations for AI UX as well. We are starting to see “who can see what my AI is doing” as a competitive feature rather than an afterthought — which feels like the right direction.
4. An Android App That Tells You When Meta’s Smart Glasses Are Nearby
Hide from Meta's spyglasses with this new Android app — The Register
Yves Jeanrenaud, a deputy professor at Darmstadt University of Applied Sciences, has published Nearby Glasses — a free Android app that alerts you when Ray-Ban Meta AI Glasses (or other smart eyewear) are in your vicinity. It works by scanning Bluetooth Low Energy (BLE) advertising frames for manufacturer company identifiers that smart glasses broadcast continuously. As Jeanrenaud explains in the project’s GitHub repo: “This app notifies you when smart glasses are nearby. It uses company [identifiers] in the Bluetooth data sent out by these [devices].”
The technical detail matters: even though BLE devices randomise their MAC addresses and service UUIDs to prevent tracking, manufacturer company IDs in BLE advertising frames are mandatory and immutable — they cannot be changed. That is the gap Jeanrenaud is exploiting. He is candid about the limitations: there will be false positives from other Meta hardware (Quest headsets, for instance), so the repo carries a prominent warning not to confront or harass anyone based solely on the app’s output.
The broader context is uncomfortable. Meta’s own LED indicator — the feature it points to as proof of transparency — can be disabled with a simple hardware mod (there are YouTube tutorials showing how). Meta reportedly has plans to add facial recognition to a future generation of the glasses. And The Register notes a cluster of recent incidents: a California judge rebuked Zuckerberg’s legal team for wearing Ray-Ban Meta glasses in court, and there have been documented cases of so-called “manfluencers” using them to covertly record women in public.
Why This Matters
This is one of those stories where the technology is almost incidental. Smart glasses sit in a genuinely awkward legal and social space: it is generally legal to record in public, but 11 US states require two-party consent for audio, and recording that involves facial recognition or constitutes harassment or stalking can quickly cross into illegality. What Jeanrenaud has built is not a perfect solution — he says so himself — but it is the kind of grassroots counter-tool that tends to appear when institutions have not yet caught up with a new surveillance surface. The fact that it needs to exist at all is the signal worth paying attention to.
5. Google Announces Nano Banana 2
Google published the Nano Banana 2 release on its official AI blog, introducing a new iteration in the Nano Banana image generation line. Officially called Gemini 3.1 Flash Image, it combines the quality of the premium Nano Banana Pro with the speed of Gemini Flash. Key capabilities include image generation at up to 4K resolution, character consistency across up to five subjects in a single workflow, precise text rendering and translation within images, and real-time grounding via web search — meaning the model can pull up-to-date information while generating. API pricing drops roughly 50% compared to Nano Banana Pro at 1K resolution, and the model is rolling out as the default across the Gemini app, Google Search AI Mode, Lens, Flow, and Google Ads in 141 countries on day one.
Why This Matters
What caught my attention this week is not just the model itself — it is how fast it moved into third-party products. According to the AI News digest from Latent Space, Nano Banana 2 appeared in Perplexity Computer on the same day it launched — a day-zero integration that connects directly with Signal 6 below. The ecosystem is no longer waiting for models to mature before building on them — new capabilities are being wired into products within hours of release. For builders, this raises the bar: your integration timeline is now measured in days, not sprint cycles.
6. Perplexity Launches Computer for Multi-Agent Task Execution
Perplexity's new tool deploys teams of AI agents
PCWorld’s headline calls it directly: “Perplexity Computer is agentic AI like OpenClaw but safer.” Perplexity has launched Computer, a multi-agent digital worker that takes a high-level goal — build a dashboard, plan a marketing campaign, create an Android app — decomposes it into subtasks, and delegates each to the model best suited for that job. The core reasoning engine runs on Claude Opus 4.6, with Gemini handling deep research, Nano Banana 2 generating images (integrated on day one — see Signal 5 above), Veo 3.1 for video, Grok for lightweight speed tasks, and ChatGPT 5.2 for long-context recall. Currently available to Perplexity Max subscribers ($200/month), with Pro and Enterprise access expected to follow.
The key architectural distinction from OpenClaw — and the reason PCWorld frames it as a “safer” rival — is that everything runs in the cloud, in isolated compute environments with a real filesystem, browser, and tool integrations, but with no access to your local machine. There is no .env file sitting next to your SSH keys.
Why This Matters
If you have read my post on OpenClaw’s architecture and risks, you will recognise this immediately. The five failure modes I walked through there — accidental mass emails, cascade deletes, Slack impersonation, credential harvesting, the relationship grenade — all share a common root: local agents inherit far too much access from the moment you install them. Perplexity Computer’s cloud-first, isolated-environment approach is a direct architectural response to that problem. You lose some flexibility (no local file access, no LAN visibility) but you gain a dramatically smaller blast radius when something goes wrong. The interesting open question is oversight: a system designed to run autonomously for hours or months still needs meaningful human checkpoints. The Nano Banana 2 day-zero integration is one small signal of how fast the model layer beneath these agents is moving — which makes that oversight question more urgent, not less.
The Agent and the Atom
Two stories this week sit slightly apart from the usual model-and-app roundup, but they deserve a section of their own. Together they trace the same arc from opposite ends: AI is getting better at using our computers, and simultaneously forcing a rewrite of global energy policy to keep the lights on.
Anthropic Acquires Vercept to Advance Claude’s Computer Use
Anthropic acquires Vercept to advance Claude's computer use capabilities
Anthropic acquires Vercept in early exit for one of Seattle's standout AI startups — GeekWire
On Wednesday (Feb 25), Anthropic announced it has acquired Vercept, a Seattle-based startup founded by alumni of the Allen Institute for AI. Vercept was built around a specific thesis: making AI genuinely useful for complex tasks requires solving hard perception and interaction problems — in other words, teaching AI to see software interfaces and act within them the way a human at a keyboard would, rather than relying on back-end APIs. Its flagship product, Vy, was a native macOS agent that ran locally on the user’s machine — no plugins, no extra logins — and could complete multi-step tasks inside live applications by seeing and acting on whatever was on screen.
Vercept’s co-founders Kiana Ehsani, Luca Weihs, and Ross Girshick will join Anthropic. Notably, co-founder Matt Deitke had already moved to Meta’s Superintelligence Lab just before the acquisition, illustrating the fierce talent competition in agentic AI. The startup had raised approximately $50 million in total, including a $16 million seed round backed by former Google CEO Eric Schmidt, Google DeepMind chief scientist Jeff Dean, and Dropbox co-founder Arash Ferdowsi. The Vy product will shut down on March 25, giving current users 30 days to migrate to Claude’s tools. This is Anthropic’s second acquisition in three months, following the purchase of coding engine Bun in December.
In Anthropic’s announcement, the company noted that Claude’s computer use benchmark performance has jumped from under 15% on OSWorld in late 2024 — when computer use was first released — to 72.5% today with Claude Sonnet 4.6. That is a remarkable trajectory in about 15 months, and Vercept’s perception and interaction expertise is aimed squarely at pushing it further.
Why This Matters
I find this acquisition genuinely exciting, and not only for the obvious strategic reasons. It signals that “computer use” — AI that can navigate software the way a person does — is moving from interesting demo to serious engineering investment. The jump from 15% to 72.5% on OSWorld is already a dramatic shift; what is interesting is how far short of 100% it still is, and what the remaining 27.5% represents: edge cases, unexpected UI states, ambiguous instructions, the messy reality of real desktops. Vercept’s work was specifically about those hard problems. Worth noting for the broader market: UiPath’s stock dropped roughly 3.6% on the news — the market is reading this as a competitive pressure on robotic process automation as a category. That reaction tells you something.
The “Nuke-and-Cloud” Push: AI’s Energy Bill Goes Nuclear
Data Center Outlook — Sightline Climate
DOE selects TVA and Holtec to advance deployment of US Small Modular Reactors
Hyperscale data centre protests — The Ecologist
The energy bottleneck is no longer theoretical, and this week several threads converged to show how serious it has become.
The grid is no longer keeping up. According to Sightline Climate’s Data Centre Outlook, 30–50% of capacity slated for 2026 is unlikely to come online before year’s end, primarily because power infrastructure is the binding constraint. Hyperscalers are increasingly bypassing the grid entirely for their largest AI training campuses. While grid-connected projects remain the most common by count, on-site and hybrid power approaches now account for nearly half of announced capacity — a remarkable imbalance driven by a small number of gigascale, grid-independent campuses. Google’s acquisition of Intersect Power’s 10.8 GW pipeline and Amazon’s direct investments in solar and storage are examples of this shift: rather than waiting for utilities, hyperscalers are buying their way to the power source.
Nuclear is moving from strategic ambition to funded project. This week the US Department of Energy awarded up to $800 million in cost-shared funding to two teams — TVA (deploying a GE Vernova Hitachi BWRX-300 at Clinch River in Tennessee) and Holtec (deploying two SMR-300 units at the Palisades site in Michigan) — to advance the first commercial Small Modular Reactor deployments in the US, targeting early 2030s operation. Secretary of Energy Chris Wright described the programme explicitly as infrastructure for “data centers and AI growth.” Meanwhile, Pennsylvania’s HB 2017, which gives state regulators authority to set lower fees for SMR and micro-reactor sites, cleared committee — one of over 350 pieces of nuclear-related legislation currently active across 46 US states.
I should be honest about the timeline here: no SMR is yet operational in the US for commercial power generation, and industry estimates put the first realistic data-centre-powering deployments at 2028–2030 at the earliest. Oracle’s much-discussed plan for a gigawatt-scale data centre powered by three SMRs remains in the planning and permitting phase, with no confirmed location or construction date. This is a 10-year bet, not a 2026 solution — and it is worth being clear-eyed about that gap.
And this week, people pushed back. Starting today (February 27), environmental charity Global Action Plan is coordinating two days of nationwide protests across the UK against the “unchecked expansion” of hyperscale AI data centres, including a “March Against The Machines” outside OpenAI’s London offices on Saturday. The numbers make the tension concrete: according to the UK energy regulator, 140 data centres have signalled they want to connect to the grid, with a combined potential power demand of 50 gigawatts — higher than UK peak electricity demand of 45 GW recorded just on February 11.
Why This Matters
The two halves of this section are deliberately placed together because they describe the same pressure from opposite angles. On the one hand, AI’s software capabilities are advancing rapidly — computer use, agentic workflows, autonomous coding — and each step up in capability comes with a step-up in compute demand. On the other side, the physical infrastructure required to support that compute is colliding with grid capacity limits, community resistance, and the practical realities of nuclear build timelines. The $710 billion capex figure from Section 3 is not abstract: it is translating into planning applications, grid connection requests, protests, and legislation, right now, in real places. The “AI as a physical infrastructure challenge” framing feels increasingly accurate.
Closing Reflection
I keep returning to a single image from this week: an AI agent that can now operate a computer with 72.5% reliability, trained in a data centre that may eventually be powered by a nuclear reactor not yet built — while the people living near that data centre march in protest, and the rest of us quietly discover that our next laptop will cost more because of it.
That is not a contradiction. It is a description of a transition. The software is moving at the pace software moves: fast, iterative, compounding week on week. The infrastructure it depends on moves at the pace of silicon fabs, power grids, planning permissions, and political will — which is to say, slowly, expensively, and with significant friction.
This week gave us glimpses of both. Vercept’s team joins Anthropic to close the last 27.5% gap in computer use. Vinext rewrites a framework in a week. Nano Banana 2 ships and is integrated the same day. Meanwhile, $710 billion in capex is causing a memory shortage that reaches into your pocket. SMR funding is approved, but the first reactor is a decade away. Protesters gather outside a London office asking who consented to all of this.
None of these threads resolves this week. But they are all part of the same story: AI is no longer just a software problem. It has become a materials problem, an energy problem, a planning problem, and — with today’s protests in mind — a social contract problem too.
Which of those problems do you think bites hardest first?
