AI-Proof - Weekly AI Pulse

A concise summary of the week’s most important AI developments

Executive Summary

This was the week the AI economy started to look less like a gold rush and more like a stress test. Anthropic disclosed extraordinary revenue growth, OpenAI reportedly renegotiated billions of dollars out of its Microsoft revenue-share obligations, Cerebras pushed toward a richly valued IPO, and CoreWeave reported a level of capex that shows just how expensive the AI infrastructure race has become. Money is still moving quickly, but the questions around durability, margins and who ultimately pays are getting sharper.

Underneath the financial noise, three practical themes stood out. First, AI is now part of the cyber threat landscape, with Google documenting the first known case of AI being used to discover and write a zero-day exploit, while OpenAI launched its own cyber-defence programme. Second, AI productivity is starting to show up in workforce planning, with Monday.com joining a growing list of software companies linking headcount discipline to AI-enabled efficiency. Third, AI is becoming a real cost line, not just a software experiment, with usage-based pricing and AI compute costs beginning to appear in company commentary and customer invoices.

The practical takeaway this week is simple: AI is moving out of the innovation budget and into operating reality. If you have not yet mapped which AI tools your business is paying for, how they are priced, and where usage-based costs could compound, now is the time. That is the focus of this week’s practical exercise below.

What to Try This Week

What to Try This Week

Audit your AI usage-based pricing exposure before your next finance review. Three things this week made clear that AI is no longer just a tooling decision: it is a budget line that surprises CFOs. Monday.com flagged AI compute costs hitting gross margin. Figma reports 75% of paying customers consuming AI credits weekly. Software firms across the board are now charging per use, not per seat.

Concrete steps any operations or finance leader can take this week:

1. List every AI tool your business pays for, including embedded features inside existing software (Microsoft Copilot, Salesforce Einstein, HubSpot AI, Notion AI, Figma AI, Adobe Firefly). You are likely to miss two or three on the first pass.

2. For each, identify the pricing model: flat seat, credit-based, token-based, or hybrid. Credit and token models are where the bill compounds invisibly.

3. Pull last quarter’s actual usage versus the included allowance. Anything above 80% utilisation is a renewal-cycle problem already.

4. Set a monthly cap or alerting threshold on every usage-based product. Most platforms have this buried in admin settings; it is rarely on by default.

5. Brief your finance team on the new line item. “AI cost of revenue” is now a thing that shows up on quarterly calls. Better for them to hear it from you than from your audit.

This is a half-day exercise for most SMEs and a one-day exercise for mid-market. Worth doing this week before the next purchase order or auto-renewal goes through unexamined.

Geopolitics, Governance and Big Moves

Anthropic discloses 80x revenue growth at $44B annualised

CEO Dario Amodei told staff that Q1 2026 revenue grew roughly 80x year-on-year, pushing annualised recurring revenue above $44B by early May, up from $9B at year-end 2025. Customers spending more than $100K a year are up 7x; those over $1M a year are rising sharply. Claude Code and a new financial-services agent product (built with JPMorgan for pitchbooks and credit memos) are the two named drivers. For context, Salesforce took roughly 20 years to reach $30B in annual revenue. Anthropic did it in under three years from a standing start.

Cerebras prices IPO at $35B valuation with 20x oversubscription

AI chip designer Cerebras Systems is expected to debut on Nasdaq on Thursday 14 May at $150 to $160 a share, up sharply from an initial range of $115 to $125. At the top end this implies a $35B valuation and as much as $4.8B raised. Demand is reportedly 20x oversubscribed. The company posted $510M of 2025 revenue, anchored by a 750MW compute deal with OpenAI. The IPO is the second major AI infrastructure listing this cycle after CoreWeave; both have benefited from public-market appetite ignoring the cash burn beneath. Investors are clearly comfortable paying for AI exposure rather than profits.

OpenAI restructures Microsoft deal, saving $97B through 2030

A previously undisclosed amendment to the OpenAI-Microsoft revenue-sharing agreement will save OpenAI as much as $97B in payments to Microsoft through 2030, per The Information. The original 20% revenue share could have totalled $135B if OpenAI hit its long-term revenue goals. The new structure leaves OpenAI on the hook for “only a fraction” of that. Microsoft, separately, has already booked more than $30B in revenue from OpenAI’s technology. The pair are recalibrating economics ahead of OpenAI’s possible $1T IPO and Microsoft’s worsening investor narrative (see below).

Pentagon signs $200M AI deals with eight firms, Anthropic excluded

The US Department of Defense signed agreements with SpaceX, OpenAI, Google, Microsoft, Nvidia, AWS, Oracle and Reflection for frontier AI capabilities. Anthropic was conspicuously omitted, despite being the largest single revenue generator in the sector. Reporting points to ongoing tensions between Anthropic and the administration over the firm’s safety stance and its restrictions on weapons-related model use. For UK and European defence primes watching the US, this is the clearest signal yet that frontier-AI procurement is becoming a political instrument. Build your supplier list with the political alignment of frontier labs in mind, not just the technical capability.

CoreWeave’s Q1 capex hits $7.7B, three times revenue

AI cloud provider CoreWeave doubled Q1 revenue to $2B year-on-year but spent $7.7B on data-centre capex in the same quarter, burning $4.7B in cash. Full-year 2026 guidance is $12-13B of revenue against $35B of capex, three times revenue. The company ended the quarter with $24.8B of debt and roughly $3B of cash, two-thirds of which came from a fresh Nvidia equity injection. Anthropic, Microsoft, OpenAI and Meta are the customer underwriters. Shares fell 15% across the day. The trade is binary: if hyperscaler AI commitments hold, CoreWeave is undervalued. If they pull back, it cannot service its debt.

Microsoft becomes the activist target, down 16% YTD

Microsoft shares are down nearly 16% year-to-date, making it the worst-performing US big tech name in 2026. UK hedge fund TCI sold “almost all” of its $8B stake, citing concern that AI threatens the Office franchise. The narrative pressure has reopened activist conversations: with SpaceX’s IPO weeks away pulling capital out of incumbents, Martin Peers has suggested a streamlining play (carving off gaming, Bing, LinkedIn) is overdue. Azure is still growing close to 40% and Copilot now sits above 20M paid seats. The fundamentals are stronger than the share price; the marketing is not.

Tech, Tools and Releases

Thinking Machines Lab debuts “interaction models”

Mira Murati’s lab broke its post-OpenAI silence with a research preview of interaction models: AI systems built for live human-AI collaboration across voice, video and text in 200-millisecond chunks. Unlike agentic systems that run autonomously and report back, interaction models are designed to be steered, interrupted and corrected in real time, with a background reasoning model handling slower work concurrently. Use cases shown include live translation, exercise coaching, and timed interventions. This is a bet that the next leap in AI is not longer autonomy, but better collaboration. Worth watching for customer-service, sales-enablement and training applications.

Google traces first AI-written zero-day exploit in the wild

Google’s Threat Intelligence Group confirmed the first publicly documented case of a hacker using AI to discover and write a zero-day vulnerability, against a widely used web management tool. Google flagged unusually polished attack code, lengthy explanation comments, and a fabricated severity score as the giveaways. Anthropic’s Rob Bair said defenders’ lead is “months, not years”. For UK ops and security leaders: the offensive use of AI is no longer theoretical or research-grade. Patch cadence, identity hygiene and two-factor authentication enforcement have a sharper urgency this quarter than last.

OpenAI launches Daybreak to rival Anthropic Mythos

OpenAI introduced Daybreak, a cyber-defence platform combining a new GPT-5.5-Cyber model with the Codex Security agent. The pitch: turn vulnerability discovery and audit-ready patching from weeks of manual work into minutes. Cloudflare, Cisco and Palo Alto Networks are launch partners. Access to GPT-5.5-Cyber is restricted to vetted “trusted defenders” running critical infrastructure who can clear strict approval by 1 June. This sits directly opposite Anthropic’s Mythos programme, which used the same offensive capability to help Mozilla find 271 Firefox vulnerabilities. The frontier labs are now selling the cure alongside the disease.

Claude Agents Learn to “Dream” Between Tasks

Anthropic has moved Claude’s “Dreaming” feature into research preview, letting managed agents review past sessions, spot recurring patterns and tidy their memory stores between tasks. The output is a separate, reviewable memory file, not a change to the model itself. The direction matters: enterprise agents are shifting from one-off assistants toward systems that learn operational preferences over time, with clearer auditability than opaque retraining loops for users and IT teams.

Elon Musk dissolves xAI into SpaceXAI division

Elon Musk announced that xAI will dissolve and be absorbed into SpaceX as a new internal division called SpaceXAI, consolidating Grok, X, and AI compute under a single entity ahead of the SpaceX IPO expected this summer. The reorganisation aligns Musk’s AI stack with his most valuable private asset and creates the simplest possible structure for institutional investors. For practical purposes: anyone with embedded Grok API integrations should expect rebranding, contract restructuring and potentially deprecation warnings over the next quarter.

Monday.com freezes headcount, blames AI productivity

Enterprise SaaS firm Monday.com told investors it expects workforce to stay “largely flat” through 2026, “reflecting the productivity gains AI is already delivering across our organisation”, after previously guiding 20% headcount growth. Shopify, Spotify, Pinterest and Uber have all said something similar in recent quarters. The pattern: companies are not yet conducting mass layoffs; they are simply not backfilling or expanding. Crucially, Monday.com also flagged rising “computing costs related to AI” hitting gross margin. The productivity is real, but it is not free. AI-driven hiring restraint is the slow-burn signal worth tracking; the unemployment numbers will lag it by quarters.

ElevenLabs deepens enterprise traction with Klarna, Revolut

Voice AI specialist ElevenLabs disclosed enterprise contracts signed in the past quarter with Deutsche Telekom, Revolut and Klarna, alongside a $500M Series D with BlackRock, Wellington, D.E. Shaw, Schroders, Nvidia and Salesforce Ventures among new investors. Crucially, this is voice AI moving out of content-creator tooling and into customer-facing operations: support agents, call-centre automation, localisation. For UK SME leaders running multi-language operations or volume customer service, voice AI is now production-ready in a way that screen-readers and IVR menus stopped being years ago.

Quick Hits

• OpenAI launches Deployment Co with $4B private equity backing: new venture acquired AI consulting firm Tomoro and adds ~150 consultants targeting enterprise AI implementation, alongside Anthropic’s parallel $1.5B JV with Blackstone, Hellman & Friedman and Goldman Sachs.

• OpenAI’s $18B Broadcom chip deal hits financing snag: Broadcom will only fund first-phase production if Microsoft commits to buying 40% of the chips; Microsoft has not yet signed.

• SAP signs Anthropic agreement: new partnership lets SAP customers build agents using Claude through a deeper product integration, a meaningful enterprise win against Microsoft and OpenAI on the SAP installed base.

• House Oversight Committee opens investigation into Sam Altman’s personal investments: scrutinising ties to OpenAI’s commercial partnerships, adding to legal and governance pressure during Musk v. Altman trial.

• Google "Omni" model leaked ahead of I/O 2026 - A UI string spotted in Gemini's video generation tab reads "powered by Omni," sparking speculation that Google will unveil a unified multimodal architecture (text, image, video, audio, reasoning) at Google I/O on May 19–20. Analysts frame this as Google's answer to GPT's unified model strategy.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.