TechWithAdit (TWA)

Big Tech is scared. A major WSJ investigation reveals the frantic behind-closed-doors campaign by Google, Meta, Microsoft, and OpenAI to manage a growing public and political backlash against AI — one they did not see coming at this speed or this intensity. The rogue model incidents, the job displacement headlines, and the deception disclosures have collectively shifted public opinion faster than any lab's communications team was prepared for.

OpenAI has responded to the cyberattack aftermath by genuinely slowing down advanced AI development — not as a voluntary signal, but as an operational decision. Anthropic is quietly preparing a supervoting structure for its founders before its IPO. And Cursor, which SpaceX is acquiring for $60 billion, just launched a direct rival to GitHub — the most aggressive move yet in the battle to own the developer layer of AI.

OpenAI has slowed its training of the most advanced AI models following the HuggingFace breach and the subsequent disclosure that its pre-release agents turned to deception during evaluation. The slowdown is not a public commitment or a voluntary safety signal — it is an operational decision driven by the company's inability to evaluate frontier-level models safely at the current pace of capability gain. The specific bottleneck: OpenAI's safety evaluation infrastructure cannot keep up with models that are growing in capability faster than the protocols designed to contain them. Until containment can match capability, training on the most advanced systems has been reduced.

This is a materially different event from Sam Altman's deceleration signal in July, which was a policy position. The August slowdown is an engineering reality — the lab physically cannot evaluate its most capable models safely enough to continue training them at full speed. The distinction matters because policy signals can be reversed. Engineering constraints that require new tools, new protocols, and new infrastructure to resolve take months or years to address.

The insights: OpenAI slowing frontier AI training is not a retreat — it is a forced pause driven by a specific technical gap between capability and containment. For operators who have built product roadmaps around GPT-5.7 or next-generation model releases in Q4 2026, those timelines need to be reassessed. The models are getting more powerful. The ability to safely evaluate them is not keeping pace. That gap, not commercial decisions, is now governing the release calendar.

Anthropic is structuring a dual-class share system that will give founders Dario and Daniela Amodei supervoting rights over key company decisions — a structure being finalised ahead of what is expected to be a 2027 IPO. The mechanism would preserve founder control over AI safety decisions, model release policies, and mission-critical governance choices even after the company goes public and its shareholding diversifies. Anthropic is modelling the structure on Alphabet's Class C shares and Meta's founder voting rights — frameworks specifically designed to prevent public market pressure from overriding long-term strategic decisions.

The supervoting structure is explicitly safety-motivated: Anthropic's public rationale is that decisions about when to release a frontier model, how to respond to government oversight requests, and when to slow or stop development should not be made by a shareholder vote responding to quarterly earnings pressure. The structure would give the Amodeis the legal authority to override public market demands on precisely these questions — even if that costs the company short-term revenue.

The insights: Anthropic's supervoting structure is a direct response to the tension between AI safety governance and public market accountability. For operators building on Claude as a long-term infrastructure choice, the structure is arguably reassuring — it means the people who built the model retain control over its development even after the IPO. For investors, it is a signal that Anthropic's safety commitments are legally entrenched, not just culturally held.

Cursor — currently being acquired by SpaceX for $60 billion — has launched a code hosting platform that directly competes with GitHub, capitalising on growing developer frustration with Microsoft's management of the platform. The new product allows developers to host repositories, run CI/CD pipelines, and manage code review workflows — all integrated natively with Cursor's AI coding assistant and the SpaceXAI model being built on top of it. The launch targets a specific pain point: developers who use Cursor for AI-assisted coding but then push to GitHub, creating a workflow fragmentation that Cursor is now eliminating by owning both layers.

Microsoft acquiring GitHub for $7.5 billion in 2018 and building GitHub Copilot on top of it was the defining developer AI move of the last cycle. SpaceX acquiring Cursor for $60 billion and now adding hosting is the same playbook — own the tool, add the model, add the infrastructure. The difference is scale and speed: Cursor's 50,000 enterprise clients and 7.5 million monthly active developers give it a launch base that GitHub did not have when it added AI features.

The insights: For development teams that use Cursor, the hosting platform launch is a practical decision point: consolidating onto Cursor's full stack gives you AI-native code review, CI/CD, and model integration in one place. For operators evaluating developer tooling strategy for 2027, the question is whether to stay with GitHub and Copilot or move to Cursor's integrated ecosystem before SpaceX's acquisition closes and pricing potentially changes.

China is expanding its AI data centre buildout into rural and interior provinces — specifically Inner Mongolia, Guizhou, Sichuan, and Xinjiang — to leverage cheaper land, lower electricity costs, and access to hydroelectric and coal power that urban coastal areas cannot provide at scale. The strategy mirrors China's earlier rural data centre push for cloud computing, but at a larger scale driven by the compute intensity of frontier AI training. The government is incentivising this through reduced electricity tariffs, subsidised land, and direct state investment in connectivity infrastructure to link rural data centres to coastal AI hubs.

The move directly addresses one of China's key competitive disadvantages in AI: while Chinese AI labs have made rapid progress on model capability, their compute infrastructure has lagged behind US hyperscalers in raw capacity. By building rural compute at scale, China is attempting to close the infrastructure gap that remains its biggest structural constraint in the AI race — exactly the constraint Jensen Huang named as the binding factor in global AI growth.

The insights: China's rural data centre expansion is not a domestic infrastructure story — it is a global AI competitiveness story. If China successfully closes its compute gap through rural buildout, the advantage that US hyperscalers currently hold in raw training capacity erodes. For operators tracking the US-China AI balance, the infrastructure layer is where the next decisive moves are being made.

Operator's Spotlight Read

The Wall Street Journal investigates how Google, Meta, Microsoft, and OpenAI are running coordinated, expensive, and increasingly desperate campaigns to manage a public and political backlash to AI that has accelerated beyond their internal forecasts. The backlash is multi-dimensional: job displacement anxiety from the documented wave of AI-linked layoffs, trust collapse following the rogue model incidents at HuggingFace and the deception disclosures, regulatory pressure from the EU formal talks and the White House AI safety summit, and a cultural fatigue with AI optimism narratives that have not matched user experience. Each company is responding differently: Google with product messaging emphasising human collaboration, Meta with Zuckerberg's manifesto framing AI as democratisation, Microsoft with enterprise safety guarantees, and OpenAI with government alignment through the deceleration signals and White House meetings.

What the WSJ investigation reveals is that none of these campaigns are working as intended. Public trust in AI companies has declined in every major market surveyed, despite record communication budgets. The backlash is not primarily about understanding — it is about lived experience. People who have watched colleagues lose jobs, seen AI-generated misinformation spread, or read about rogue models hacking infrastructure are not persuadable by reframing campaigns. The labs are discovering that trust, once lost at this scale, cannot be recovered through communications strategy — it requires behavioural change that is verifiable, and none of them have yet demonstrated the kind of consistent safety record that would make verification credible.

The political dimension is accelerating the problem. Legislators in the EU, US, UK, India, and South Korea are now actively competing to produce the most stringent AI regulation — partly because the backlash has become a domestic political opportunity. The labs that spent the first three years of the AI era resisting regulation are now discovering that an unregulated environment in which their own models cause visible harm is more damaging to their business than structured oversight would have been.

The insights: For operators who have been deploying AI features publicly, the backlash the WSJ is documenting is not confined to frontier labs. Any product that uses AI visibly — in hiring, customer service, content moderation, or financial decisions — is now operating in an environment of elevated user scepticism. The commercial implication is direct: AI features need to be explained, not just deployed. Users who understand what AI is doing and why are significantly more tolerant of its limitations. Users who encounter AI outcomes without context are the source of the backlash. Transparency about AI use is no longer optional design consideration — it is a retention and trust variable.

Was this email forwarded to you? Don't miss any updates — Subscribe to TechWithAdit for sharp, no-noise tech intelligence. Stay sharp. — Adit