The AI Breakdown

This week: the dependence problem running underneath the AI boom. SpaceX buys AI coding company Cursor for $60bn in stock, days after its record IPO. Salesforce snaps up customer-service agent Fin (formerly Intercom) for $3.6bn. ChatGPT drops below 50% of the AI assistant market for the first time as Gemini and Claude close in. Two of Google's top AI researchers depart for rivals (Transformer co-inventor Noam Shazeer to OpenAI and Nobel laureate John Jumper to Anthropic), and Alphabet has its worst day in about a year. Anthropic's Fable 5 and Mythos 5 are still offline ten days into a US export-control order. America's energy regulator, FERC, votes to fast-track AI data centres onto the grid while shielding household bills. A new "agentjacking" attack hijacks AI coding agents through fake error reports. And PwC's AI Jobs Barometer finds the most AI-exposed firms are growing wages and headcount, not cutting. Plus a fast look at the model race: GPT-5.6, Gemini 3.5 Pro and China's open-weight GLM-5.2. If you find this useful, follow the show and leave a rating.

A big week for assistant strategy, pulling in two directions. Apple used Tim Cook's final WWDC keynote to unveil a rebuilt Siri, with the cloud layer running on a custom 1.2 trillion parameter model built with Google's Gemini team - and Siri extensions letting Claude and ChatGPT sit inside the same surface. Microsoft went the other way at Build 2026, pulling its AI stack in-house: seven MAI models, an always-on agent called Scout, and a native GitHub Copilot desktop app. Then the money. The US government is reportedly discussing a direct equity stake in OpenAI, while Bernie Sanders floats a one-time stock tax on the big labs. SpaceX kicked off the largest IPO roadshow on record, now priced as much on AI compute demand as on rockets. And Meta launched its Business Agent across WhatsApp, Instagram and Messenger, betting small firms never leave the thread. Plus the quick hits: ChatGPT's biggest memory update since launch, a new Lockdown Mode, Alibaba's Qwen 3.7 Max undercutting Western labs on price, and OpenAI's life-sciences model GPT-Rosalind.

This week on The AI Breakdown, we look at the moment AI starts to meet the public markets. OpenAI has reportedly moved toward a confidential IPO filing, Anthropic is showing signs of operating profit while leaning heavily on massive compute deals, and Google used I/O to remind everyone that distribution may matter as much as model quality. We also cover Anthropic’s SpaceX compute dependence, Nvidia’s huge AI infrastructure numbers, Meta’s messy AI reorganisation, and Andrej Karpathy’s move to Anthropic as the frontier AI talent war heats up.

This week, the AI infrastructure wars went into overdrive. Anthropic and OpenAI both announced massive deployment vehicles - Anthropic's $1.5 billion joint venture with Blackstone, Goldman Sachs and Hellman & Friedman, and OpenAI's $4 billion Deployment Company plus the acquisition of UK consultancy Tomoro - signalling a direct shot at Accenture, Deloitte and the rest of the Big Consulting machine. The message? The bottleneck is no longer the model, it's getting it into production. Then there's the deal nobody saw coming. Anthropic has leased the entire Colossus 1 data centre from Elon Musk's SpaceX - 300+ megawatts, 220,000 Nvidia GPUs - after Dario Amodei revealed Claude grew 80x in Q1 against an internal plan of 10x. Rate limits are easing, Opus token caps are jumping tenfold, and Musk has gone from calling Anthropic evil to becoming its biggest landlord. Plus: Microsoft Agent 365 hits general availability with a serious control-plane play, AWS launches Bedrock AgentCore Payments so agents can transact in stablecoins, SAP drops €1 billion on tabular foundation model startup Prior Labs, Sierra raises $950 million at a $15 billion valuation, OpenAI opens GPT-5.5-Cyber to European defenders while Anthropic keeps Mythos walled off, Alibaba bakes Qwen into Taobao across 4 billion products, and Cloudflare cuts 1,100 jobs citing internal AI productivity gains - only for the market to punish the stock anyway.

Google and Amazon just put up to $65 billion and 10 gigawatts of compute behind Anthropic in five days, and that tells you the AI market is no longer just about models. It is about cloud lock-in, silicon validation, procurement confidence and who gets to become the default infrastructure layer for enterprise AI. This weekly roundup unpacks Google’s planned $40 billion Anthropic investment, the five-gigawatt TPU commitment, and why it removes one of Claude’s biggest objections in Fortune 500 buying cycles. You also get the AWS side of the story: Amazon’s expanded $25 billion backing, more than $100 billion in Anthropic spend on AWS over a decade, and the bigger point that Anthropic is becoming an anchor tenant in the cloud wars, not merely a well-funded lab. From there, the focus shifts to OpenAI’s GPT-5.5 release, with its one million token context window, multi-step agentic workflows and the growing sense that frontier models now update on software cadence rather than annual launch cycles. You also get the labour signal many leaders have been waiting for, as Meta and Microsoft cut a combined 17,000 roles while ramping AI capex, making workforce strategy and AI investment impossible to treat as separate conversations. The episode closes on Google replacing Vertex AI with the Gemini Enterprise Agent Platform, plus quick takes on Cursor, Merck and Google Cloud, Adobe’s CX Enterprise rebrand, and China blocking Meta’s Manus deal.

Meta's Muse Spark may have just shown that in AI, distribution matters more than benchmark bragging rights. This week I break down why Meta's new multimodal model matters beyond the rankings, why its HealthBench Hard score stands out, and why pushing a free model into Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta glasses could be a stronger commercial move than simply topping the leaderboard. I also dig into Anthropic's 3.5-gigawatt compute deal with Google Cloud and Broadcom, its reported $30 billion annualised revenue run rate, and what more than 1,000 enterprise customers each spending over $1 million a year says about where the real enterprise AI race is heading. From there, the episode turns to the anti-distillation coalition between OpenAI, Anthropic, and Google - including the 16 million exchanges and 24,000 fraudulent accounts Anthropic says were tied to Chinese firms such as DeepSeek, Moonshot AI, and MiniMax. Elsewhere: Google's NotebookLM and Gemini integration, BCG's warning that 50–55% of US jobs will be reshaped by AI rather than simply replaced, PwC's finding that 74% of AI's economic gains are going to just 20% of firms, Anthropic's new $0.08-per-hour managed agent infrastructure, Project Glasswing putting a restricted frontier model in the hands of Apple, Microsoft, and AWS for cybersecurity, Upwork inside ChatGPT, and why NIST's quiet work on AI agent standards in healthcare, finance, and education matters more than it sounds.

A federal judge just told the Pentagon it cannot punish Anthropic for insisting on AI guardrails. Judge Rita Lin's injunction was unusually blunt, and it may matter more than any model launch this week. This episode covers what that ruling means if AI vendors can now push back on autonomous weapons and mass surveillance clauses without being frozen out of public-sector work, plus the immediate business stakes for defence buyers already using Claude. Then there's OpenAI's very different week. Sora is being shut down after reportedly burning around $1 million a day in inference costs - with peak daily costs reaching $15 million - against just $2.1 million in lifetime revenue. That retreat connects to a wider pivot: enterprise capability, data centres, IPO optics, and the upcoming Spud model. The question worth asking is what happens when cost pressure and thinner safety oversight arrive at the same time. Also in this episode: Shopify making AI commerce the default across 5.6 million stores. A Duke and Federal Reserve-backed CFO survey projecting over half a million AI-related job cuts in 2026. Google's TurboQuant breakthrough, which could cut large language model memory needs by 6x and challenge assumptions about endless GPU demand. Google's free Gemini personalisation push. And Mistral's €830 million infrastructure bet near Paris.

This week I'm doing something a bit different. Instead of reacting to a product launch or a funding round, I'm digging into the Stanford Emerging Technology Review 2026, specifically its AI chapter. It's co-chaired by Condoleezza Rice, Jennifer Widom, and Amy Zegart, produced across Stanford's School of Engineering, the Hoover Institution, and the Institute for Human-Centered AI, and it's deliberately trying to inform rather than advocate. I pull out the parts I think deserve serious attention from technology leaders: the gap between AI's foundational promise and its current operational fragility, the overhyped narrative around agents, the hollowing out of the public research pipeline, and the governance and legal shifts that should already be affecting decisions your organisation is making today. You can find the Stanford Emerging Technology Review 2026 here: https://setr.stanford.edu/

Anthropic’s Pentagon standoff did more than derail a $200 million contract. It turned AI ethics into a live commercial test, sent Claude to number one on Apple’s US App Store, and forced a much bigger question into the open: can trust become a business model in AI? In this episode, I break down how Anthropic rejected Pentagon language allowing Claude to be used for any lawful use, why Dario Amodei pushed for explicit limits on domestic mass surveillance and fully autonomous weapons, and how the fallout escalated into a federal ban, a supply chain risk designation, and a very public consumer backlash against ChatGPT. I then contrast that with OpenAI’s launch of GPT-5.4, where the real story is not the branding but computer use: a model that can read your screen, control mouse and keyboard inputs, and move across messy enterprise systems like a junior operator rather than a chatbot. The episode also unpacks Google’s Gemini 3.1 Flash-Lite pricing move and what commodity economics could mean for AI features, China’s 15th Five-Year Plan and its state-led push into AI, quantum and robotics, Netflix’s acquisition of Ben Affleck’s InterPositive and the rise of AI as invisible production leverage, and Meta’s $600 billion infrastructure bet with AMD. Add in AI-enabled cyberattacks on FortiGate devices, new state laws in Oregon, Utah and Vermont, and Gartner’s $2.52 trillion AI spending forecast, and this becomes a sharp 20-minute briefing on where AI strategy, policy and business reality are colliding right now.

This week I'm unpacking OpenAI's record-breaking $110 billion raise and what Amazon and NVIDIA's involvement tells us about a partner landscape that's shifting faster than most people realise. I also dig into Anthropic's $30 billion Series G, and why it's time to take that one seriously as a strategic bet. Then there's Apple quietly admitting it can't build AI fast enough, handing Siri's core logic to Google Gemini. Also, the hyperscaler spending numbers are extraordinary, and I explain why the energy and infrastructure story is just as important as what's happening at the model layer. Plus: a reality check on Microsoft Copilot's 3.3% penetration, the Snowflake and OpenAI data gravity play, Samsung's push to put Gemini on 800 million devices, and what a protest march through London's tech hub on a Saturday morning tells us about where the regulatory conversation is heading.

700 million people now use AI every week. Are we keeping up with the risks? Over 100 experts from 30+ countries just published the most comprehensive global assessment of AI risk ever produced. In this episode, I break down the International AI Safety Report 2026 — what AI can actually do today, the three categories of risk every business needs to understand, why some AI systems now behave differently when they know they're being tested, and the research that's changed how I think about my own AI use. Read the full report: https://internationalaisafetyreport.org/

This week on The AI Breakdown, we talk about OpenAI’s Frontier launch, an enterprise platform designed to help organisations build, deploy, and govern AI agents across real workflows. Anthropic fires back with Claude Opus 4.6, including a one million token context window in beta and new agent teams designed to split complex work across multiple cooperating agents, with a clear push beyond coding into everyday knowledge work like documents, spreadsheets, and presentations. We then zoom out to the money and the infrastructure. Google is introducing a Workspace add on called AI Expanded Access from March 1, 2026, signalling the shift toward paid higher tier usage. Cerebras just closed a one billion dollar Series H at about a twenty three billion valuation, as demand for compute fuels a new wave of AI hardware competition. Finally, Super Bowl LX made AI advertising feel like a cultural inflection point. Anthropic used its spot to promise Claude will remain ad free, while OpenAI ran a Codex ad built around the idea that you can just build things now. iSpot data reported by AdWeek says 23 percent of Super Bowl commercials featured AI, and Axios covered X rolling out BrandRanx to track ad conversation in real time as the game unfolded. And with the echoes of the Dot Com Super Bowl and the Crypto Bowl still fresh in marketers minds, it raises the question, will the Super Bowl burst the AI bubble?

CES 2026 quietly marked a turning point. AI stopped being the shiny feature you bolt on for a headline, and started behaving like electricity. It's just assumed. In this episode, I break down the five AI themes that defined the show: AI PCs go mainstream: Your next laptop refresh might be your biggest AI decision this year. AMD, Intel, and Qualcomm are putting serious on-device capability into enterprise hardware, and that changes where your AI runs, how your data moves, and how much control you actually have. Edge AI and the "no cloud required" wave: From real-time deepfake detection on laptops to Caterpillar embedding a voice assistant into excavators that works without connectivity, on-device AI is solving the unsexy problems: latency, offline reliability, and data control. ROI-first AI: Siemens and PepsiCo showed what "show me the numbers" looks like: 20% throughput gains and 90% of issues caught before physical changes. The pilot era is over. Physical AI: Boston Dynamics' Atlas is heading to Hyundai factories by 2028. But the nearer-term story? Copilots for heavy machinery that upgrade the tools you already have. Trust as the bottleneck: The limiting factor isn't clever models. It's governance, guardrails, and getting your data right. If you take one thing from CES 2026: AI is becoming infrastructure. That means it's going to be boring, expensive, and absolutely worth getting right.