AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
Another busy week in the world of AI - if i was to give the week a theme it would be “Agents”.
The first shift was commercial. Ramp’s latest AI Index showed Anthropic overtaking OpenAI on paid business adoption for the first time, with Anthropic at 34.4% of paying businesses versus OpenAI at 32.3%. That does not mean OpenAI is losing the market, but it does mean the default procurement assumption has changed. Anthropic is no longer the “technical alternative”; it is becoming a mainstream enterprise option, helped by stronger business adoption, a reported $30 billion annualised run-rate and the launch of Claude for Small Business.
The second shift was infrastructure. Nvidia’s latest results suggested AI demand is still compounding rather than plateauing, while Cerebras’ IPO performance and inference-speed claims show how aggressively investors and customers are looking for alternatives to the GPU-heavy status quo. The message is not that Nvidia is about to be displaced. It is that AI infrastructure is now a strategic market in its own right, with speed, cost and availability becoming board-level questions.
The third shift was product architecture. Google I/O pushed Gemini deeper into Search, Android, Workspace, Chrome and developer tools, while Cursor, Amazon, Apple, Microsoft and Cerebras all pointed in the same direction: AI is moving (at pace) from chat windows into workflow layers. The practical implication for software vendors is uncomfortable. Lightweight AI features bolted onto legacy products will be harder to defend if larger platforms can place capable agents directly into the tools customers already use.
For business leaders, the takeaway is simple: re-run your AI vendor assumptions, test where agents could replace manual workflows, and be sceptical of product roadmaps that treat AI as a feature rather than a new operating layer.
What to Try This Week
What to Try This Week
Run a one-hour “agent exposure” review of one workflow in your business.
Pick a process that still involves people moving information between systems: preparing a sales proposal, updating a CRM, reconciling invoices, triaging support tickets, summarising customer calls, drafting board materials or checking supplier risk.
Then ask three questions:
Which steps require judgement, and which are just retrieval, formatting, checking or routing?
Which systems does the person need to access to complete the task?
If an AI agent could safely read those systems and take limited actions, what would still need human approval?
The aim is not to automate the workflow this week. The aim is to understand where the boundary sits between “AI as assistant” and “AI as operator”.
For most businesses, the opportunity will not start with replacing a whole job. It will start with removing the repetitive hand-offs inside a job. That is why this week’s Google, Anthropic, Microsoft, Cursor and Amazon announcements matter: the agent layer is moving closer to the tools your team already uses.
Geopolitics, Governance and Big Moves
Anthropic overtakes OpenAI on business AI adoption
Ramp’s April AI Index showed Anthropic taking the paid business-adoption lead from OpenAI for the first time, 34.4% against 32.3%, with Anthropic’s share rising 3.8% in April while OpenAI’s fell 2.9%. Across twelve months Anthropic quadrupled its share while OpenAI’s grew 0.3%. Ramp economist Ara Kharazian called it “a stunning reversal in the competitive market dynamics for AI model providers”. The data tracks 50,000-plus US firms’ card-and-invoice spend, so big transformation deals are not in the sample, but the trend break is unambiguous.
Cerebras Day 1 closes up 68% at a $94 billion market cap
Cerebras’ first day of public trading on Friday closed at $311.07, up 68% from the $185 IPO price set 24 hours earlier, putting the company at roughly $94 billion in market value, more than 2.3 times its $40 billion IPO valuation. The IPO raised $5.55 billion at 20 times oversubscription, making it the biggest US IPO of 2026 so far. The forward multiple is rich: Cerebras is projected to roughly quadruple revenue from $800 million this year to $3.2 billion next year against heavy OpenAI customer concentration. Investors paid the premium regardless. The public-market filter on AI-infrastructure messiness has plainly weakened.
Nvidia’s results say the AI infrastructure boom is still accelerating
Nvidia’s latest results landed like another stress test for the “AI bubble” argument. Q1 FY2027 revenue hit $81.6 billion, up 85% year on year, with data centre revenue reaching $75.2 billion. The bigger signal was forward-looking: $91 billion of Q2 guidance, a higher dividend and an $80 billion buyback. Demand for AI infrastructure is not plateauing yet; it is still compounding.
Karpathy joins Anthropic’s frontier research team
Andrej Karpathy, OpenAI co-founder and former Tesla AI lead, has joined Anthropic’s pre-training team. The move is significant because pre-training is where the core capability of frontier models is built, and Anthropic says Karpathy will work on using Claude to accelerate that research. It is another signal that Anthropic is not just winning enterprise attention; it is also attracting the research talent needed to keep pushing Claude forward.
OpenAI fights governance pressure on three fronts at once
Three escalations landed inside a single 24-hour window on Thursday. The Information’s Aaron Tilley reported OpenAI’s extraordinary decision to issue a public legal threat against Apple as both companies prepare to compete in AI hardware. A fresh court filing in the Musk v Altman trial revealed Sam Altman holds more than $2 billion in personal stakes in companies that have done business with OpenAI, fresh material for the House Oversight investigation. And Microsoft executive Michael Wetter testified that Microsoft’s commercial commitments to OpenAI will exceed $100 billion by the end of its June fiscal year. Martin Peers’ read: the “never-ending series of dramas” is now hurting OpenAI’s standing as a reliable enterprise partner.
OpenAI clears Musk lawsuit, but the governance questions remain
A California jury unanimously rejected Elon Musk’s lawsuit against OpenAI, finding he waited too long to bring the case. The ruling removes a major legal overhang as OpenAI prepares for a possible IPO. Musk plans to appeal, so the dispute is not finished, but OpenAI has won the immediate round.
Microsoft’s $100 billion-plus OpenAI exposure now publicly quantified
Wetter’s Thursday testimony nailed down a figure that had previously been triangulated. Microsoft’s commercial commitments to OpenAI will exceed $100 billion by the end of its June fiscal year: the $13 billion equity stake plus training and data-centre infrastructure. Read alongside Microsoft’s $30 billion of OpenAI-related revenue between 2023 and 2025, the relationship is recalibrating from joint venture to a highly profitable but increasingly narrow commercial partnership. With Microsoft shares down 16% year-to-date and TCI having exited its $8 billion stake, activist pressure is rising. Peers’ parallel to ValueAct’s 2013 playbook, which culminated in Steve Ballmer’s exit and Satya Nadella’s appointment, looks more plausible by the week.
Microsoft MDASH beats Anthropic’s Mythos on the CyberGym benchmark
Microsoft unveiled MDASH on Thursday: a multi-agent AI security harness chaining more than 100 specialised agents in a scan, debate and validate pipeline. One set of agents scans code for potential vulnerabilities, a separate group debates whether each finding is real and exploitable, and a final stage builds proof-of-concept attacks. In its initial deployment MDASH found 16 flaws across Windows and surpassed Anthropic’s Claude Mythos on the CyberGym benchmark. This is the first non-frontier-lab system to win that benchmark and the cleanest signal yet that multi-agent architectures with explicit debate-then-validate steps outperform monolithic frontier models on offensive-cyber tasks. Britain’s AI Safety Institute confirmed the same day that AI cyber capability is doubling every 4.7 months.
Tech, Tools and Releases
Claude for Small Business ships an SMB-ready package
Anthropic launched Claude for Small Business on Thursday, a bundle of connectors and ready-to-run workflows embedding Claude inside QuickBooks, PayPal, HubSpot, Google Workspace and Microsoft 365, plus new tools for payroll, invoices and campaigns. Anthropic was traditionally the developer-tooling specialist; pairing this with the same-day Ramp data showing Anthropic ahead of OpenAI on business adoption signals a deliberate push into the operator tier where ChatGPT had been dominant. Paid Claude plans will also receive a dedicated monthly credit for programmatic usage starting 15 June, packaging API usage into existing subscription tiers.
Google I/O turns Gemini into an agentic operating layer
Google I/O was not just another model launch. It was Google pushing Gemini deeper into the surfaces where work already happens: Search, Android, Workspace, Chrome, developer tools and, eventually, smart glasses. Gemini 3.5 Flash is being positioned for faster agentic and coding workloads, while Gemini Omni expands Google’s multimodal ambitions. The bigger signal is Antigravity and Managed Agents, which make it easier for developers to build, orchestrate and deploy agents that can use tools, run code and complete multi-step tasks. For software vendors, the warning is clear: lightweight AI features bolted onto legacy products will be harder to defend.
Adaption’s AutoScientist automates frontier-style fine-tuning
Adaption, the startup founded by former Cohere VP of Research Sara Hooker, introduced AutoScientist on Thursday: a system that automatically customises AI models for specific jobs by adjusting both training data and training settings, iterating until the model meets the goal. In internal tests AutoScientist outperformed expert-tuned models by 35% on average, with success rates jumping from 48% to 64% across finance, legal, medical and five other industries. The strategic point: only a few thousand people globally can properly fine-tune a frontier model, and nearly all of them work at the same handful of labs. If that expertise can be automated, custom per-business models become practical to build at scale.
Amazon’s Alexa for Shopping replaces Rufus and takes over the search bar
Amazon’s full Alexa for Shopping rollout details landed on Thursday across The Rundown, TLDR and Martin Peers’ Briefing. The product replaces Rufus, takes over the Amazon.com search bar, and draws on Amazon’s catalogue, reviews, delivery timing, past purchases and Alexa voice conversations to answer questions, run side-by-side comparisons, track pricing and Auto-Buy items when prices hit a target. A new Buy for Me flow handles checkouts on non-Amazon stores; Scheduled Actions auto-restocks products on a cadence. For UK retailers, the open question is what the SEO and product-catalogue playbook looks like when most shopping queries are agent-mediated.
Apple readies AI agents in the App Store ahead of WWDC
The Information’s Aaron Tilley reported on Thursday that Apple is exploring how to welcome AI agents into the App Store while preserving its privacy and security positioning. The piece described concrete policy work on letting AI agents follow Apple’s privacy rules and on preventing agents from spinning up smaller apps after the parent app has been approved. The companion announcement: iOS 27 will feature a redesigned Siri, a customisable Camera app for professionals and system-wide design changes, set for WWDC on 8 June. Read alongside the OpenAI-Apple legal flare-up the same week, Apple is opening App Store policy for AI agents while signalling its native AI rebuild will compete with whatever OpenAI ships in hardware.
Cerebras shows why AI speed matters
Cerebras says it has set a new inference-speed record for Moonshot AI’s Kimi K2.6, a huge open-weight model with around a trillion parameters.
In plain English: inference is the part where an AI model gives you an answer, and tokens are the chunks of text it generates. Cerebras reportedly reached 981 output tokens per second, meaning the model could respond far faster than the official version. That is news worthy because businesses do not just want smarter AI; they want AI that can work quickly enough to power real-time apps, agents and customer-facing tools. It also strengthens Cerebras’ argument that its wafer-scale chips offer a serious alternative architecture to traditional GPU-heavy AI infrastructure.
Cursor’s Composer 2.5 shows the cost curve for AI coding is falling
Cursor has released Composer 2.5, its upgraded in-house coding model, built on Moonshot’s open-weight Kimi K2.5 checkpoint. The claim is not just better benchmarks; it is better long-running coding behaviour at far lower cost. Cursor says pricing remains $0.50 per million input tokens and $2.50 per million output tokens, making near-frontier coding performance increasingly accessible to ordinary development teams.
Quick Hits
Anthropic CFO confirms $30 billion run-rate. Krishna Rao’s Thursday interview confirmed Anthropic at a $30 billion annualised run-rate, up from $250 million two years ago, on $75 billion raised in his tenure. WSJ separately reported investment offers arriving at $900 billion-plus valuations.
Vercel AI Gateway production trends. 200,000-plus teams now route through the Gateway, with agentic workloads carrying 59% of all token volume. Anthropic leads in spend, Google leads in volume, and open-source models are gaining traction with no lab-loyalty.
Microsoft hits $5.5 trillion through Nvidia’s market cap milestone. Nvidia became the first company ever to reach $5.5 trillion in market capitalisation, with the milestone coinciding with Jensen Huang’s arrival in China for Trump’s meetings with Xi Jinping. Nvidia H200 China-export delivery, meanwhile, remains at zero.
Cisco and LinkedIn announce layoffs into AI tailwinds. Cisco cut around 4,000 jobs, roughly 5% of its workforce, “as it shifts more spending toward AI and cybersecurity” despite 12% revenue growth. LinkedIn separately laid off 5% of its 17,500-plus global staff. Layoffs.fyi is tracking 103,000-plus tech layoffs across 2026.
ChatGPT Trusted Contact launches. OpenAI shipped a feature that lets users nominate one adult to be notified if a ChatGPT conversation suggests a potential self-harm concern. Five-step setup in Settings, rolling out worldwide within weeks. Worth raising in a team wellbeing channel.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.