Published on LinkedIn and amitabhapte.com  |  8 June 2026

This week, the companies that know AI best started saying things the rest of the market was not ready to hear. The infrastructure deals got bigger. The business model assumptions started cracking. And from San Francisco to Tokyo, the people with the most skin in the game began updating their timelines. Not upward. Downward.

When AI Starts Building Itself

More than 80% of the code now being merged into Anthropic’s own codebase is written by Claude. Not by human engineers. By Claude. In a detailed post published Thursday, the company warned that AI task-completion horizons have been doubling roughly every four months, and that recursive self-improvement. the point at which AI improves itself without human involvement. may arrive sooner than widely assumed. The post called for a coordinated, verifiable mechanism across major AI labs to slow or pause development if that threshold is crossed before society can absorb the implications.

The same week, SoftBank’s Masayoshi Son told CNBC that OpenAI’s next model is already being designed by a model. His previous superintelligence timeline was ten years, then four years. Now he says two. The contrast is striking. The company calling for a coordinated pause is the same one reporting that its own AI is already writing most of its code. The investor accelerating his timeline is the one with $65 billion committed to OpenAI. These are not contradictory positions. They are two honest readings of the same data.

My PoV: The capability curve is now steep enough that even the builders are uncertain what comes next. That is not a reason to stop. It is a reason to govern. For enterprise leaders, the practical question is not whether recursive self-improvement is real. It is whether your AI adoption roadmap has any provision for governance, verification, and escalation paths if the systems you deploy start operating outside the intent they were designed for. Most do not.

The Infrastructure Land Grab

Google has signed a $30 billion compute deal with SpaceX, paying $920 million a month to access xAI’s data centres. 110,000 Nvidia GPUs, secured from October this year through June 2029, as bridge capacity for surging Gemini Enterprise demand. This comes weeks after Anthropic disclosed it is paying SpaceX $1.25 billion a month for access to the same Colossus infrastructure. Google and Anthropic. Two competing AI labs. Both buying capacity from a third rival’s data centres. The structure of this industry is stranger than anything a traditional technology analyst would have modelled.

Microsoft, meanwhile, announced Project Solara at Build 2026, a platform designed for what it calls the agent-first device era. Reference hardware includes a smart display and a wearable smart key badge, both built on Android, both designed to execute tasks across Microsoft 365 rather than run apps. Target, CVS Health and Best Buy will pilot Solara devices in the coming months. The platform supports multiple agents, not a single dominant one, with a planned dispatcher to route work across them. The architecture of the PC is being redesigned around delegation, not interaction.

My PoV: When competitors buy compute from each other and device form factors are rebuilt around AI agents, the platform war has entered a phase that is no longer primarily about software. It is about physical infrastructure, supply agreements, and hardware ecosystems. Enterprise technology strategies that are still organised around software vendors and SaaS procurement cycles are operating on a model that is quietly becoming obsolete. The supply chain for intelligence now looks more like energy than it does like software.

The Cracks in the Business Model

A new discipline is taking hold in corporate AI spending. Model routing matches the complexity of a task to the cost of the model required to handle it. Simple queries go to cheap, fast alternatives. Hard problems go to frontier models. Glean’s CEO estimates that roughly 95% of enterprise AI usage still runs on the most expensive models, even for work cheaper models could handle. Cognition’s CEO suggests five to ten times better cost efficiency is available today on routine tasks. CFOs are paying attention. Both OpenAI and Anthropic have built their IPO-level valuations on the assumption of sustained premium-price demand. If routing takes hold, that assumption changes.

The cybersecurity earnings this week offered a different version of the same lesson. Palo Alto Networks and CrowdStrike both beat estimates, raised guidance, and cited Anthropic’s Mythos as a genuine demand inflection. Shares fell anyway, by 3% and 8% respectively. The Mythos-driven rally had already priced in a windfall that one strong quarter could not confirm. Elsewhere, OpenAI launched Lockdown Mode, an optional protection against prompt injection for users handling sensitive data. And Google announced that publishers can now opt out of AI Overviews and AI Mode entirely, without affecting their standard search ranking. Trust and consent are becoming product features. That matters.

My PoV: Model routing is not a niche technical choice. It is the beginning of enterprise AI procurement growing up. The first generation of AI spending was about access. The second will be about efficiency. For technology leaders, that means building routing logic into your AI architecture now rather than waiting for finance to force the conversation. On cybersecurity: the lesson from this week’s earnings is not that results were weak. It is that expectations, once inflated by a single model release, are hard to manage. Plan for that when you brief your leadership on what Mythos or any successor capability actually means for your security posture.

My Takeaway This Weekend

Three separate stories this week converged on the same uncomfortable truth. The people building AI are uncertain about what they are building. The infrastructure supporting AI is being traded between competitors as a strategic commodity. And the pricing models that justified trillion-dollar valuations are under pressure from the same efficiency discipline that AI was supposed to unlock everywhere else.

The leadership challenge is not to resolve that tension. It is to hold it clearly. Move forward on deployment. Build governance alongside it. Scrutinise your AI spending for efficiency. And if someone tells you superintelligence is two years away, ask what your organisation would do differently if they were right.