Two forces are reshaping the AI industry simultaneously: capital is now required at a scale that only a handful of entities can sustain, and model intellectual property is emerging as a front in the broader US-China technology conflict. Four stories from the past 48 hours make both pressures concrete.

White House accuses Moonshot AI of distilling Anthropic's Fable; Treasury threatens sanctions

On 22 July, Michael Kratsios, Director of the White House Office of Science and Technology Policy, wrote publicly that "We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model." The allegation is that Moonshot built an internal platform capable of large-scale extraction from US frontier models while rotating access methods to avoid detection. Kratsios distinguished the alleged operation from legitimate distillation, characterising it as "IP theft and industrial espionage that supports adversary military and intelligence capabilities."

Anthropic's head of public policy, Sarah Heck, supported the statement. The Treasury Department separately threatened sanctions. A further allegation: Moonshot acquired Nvidia GB300-equipped servers through Thailand, potentially in violation of US export controls.

Some independent researchers dispute the technical premise. Fable only became publicly available on 1 July, which leaves a narrow window for it to have been a primary training source for Kimi K3 at the model's documented scale. Moonshot's K3 open weights remain scheduled for 27 July. Regardless of how the technical question resolves, the incident establishes a precedent: the US government is now treating frontier model outputs the same way it treats semiconductor designs — as a national-security asset subject to enforcement action and potential sanctions.

Alphabet Q2: Google Cloud surges 82 per cent, capex guidance rises to $205 billion, free cash flow turns negative

Alphabet reported second-quarter 2026 results on 22 July. Consolidated revenue reached $119.8 billion, up 24 per cent year on year, above consensus. Google Cloud grew 82 per cent to $24.8 billion, with operating income of $8.8 billion — more than three times the year-ago figure. The cloud backlog stood at $514 billion, rising by more than $50 billion sequentially in a single quarter.

The figure that moved markets was capital expenditure. Alphabet raised its full-year 2026 capex guidance to $195-205 billion, up from $180-190 billion last quarter, with the finance chief noting the company is "still in a supply-constrained environment." Quarterly capex of $44.9 billion pushed free cash flow to -$5.9 billion — the first negative reading in recent memory for a company that ordinarily generates tens of billions annually. Approximately 60 per cent of that spend goes to servers; the remainder to data centres and networking.

On the model side: Gemini now processes 22 billion API tokens per minute (up from 16 billion last quarter), the Gemini app has 950 million monthly active users, and Gemini 4 was confirmed as being in training during the earnings call. The practical reading: even the most cash-generative company in AI is now investing at a pace that produces negative free cash flow. The constraint is construction and power, not capital or demand.

Google ships Gemini 3.6 Flash and 3.5 Flash Cyber; Gemini 4 named in training

On 21 July, Google released three models: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber — a security-focused variant designed to identify software vulnerabilities. The efficiency numbers on 3.6 Flash are operationally significant:

  • Output token usage falls 17 per cent versus 3.5 Flash — a direct cost reduction on existing workloads without a prompt change
  • Output pricing drops from $9.00 to $7.50 per million tokens
  • DeepSWE coding benchmark: 49 per cent (up from 37 per cent for 3.5 Flash)
  • OSWorld computer-use: 83.0 per cent (up from 78.4 per cent)
  • Knowledge cutoff advances from January 2025 to March 2026

Google's decision to ship three incremental models rather than wait for Gemini 3.5 Pro — which has now missed three successive release targets — suggests a deliberate choice to maintain delivery cadence ahead of flagship readiness. For teams currently running 3.5 Flash workloads, 3.6 Flash is a credible upgrade at its new price point without any context-window or integration change.

OpenAI announces Project Camellia: 3.2 GW, at least $20 billion, first self-designed campus

Also on 22 July, OpenAI announced Project Camellia, a data centre campus in Effingham County, Georgia. The facility will cover 1,400 acres, house four buildings totalling approximately 4.4 million square feet, and draw up to 3.2 gigawatts under a 25-year agreement with Georgia Power. OpenAI described the minimum capital commitment as $20 billion; Bloomberg reported the total could exceed $30 billion at full build-out.

This is the first campus OpenAI has designed and built itself, rather than leasing capacity from a hyperscaler. Power delivery runs in phases from 2028 through 2032. OpenAI will fund all grid infrastructure so that existing customers of Georgia Power bear no incremental cost. The 3.2 GW commitment represents roughly one-third of the 9,885 MW the Georgia Public Service Commission approved for new generation in December 2025, making OpenAI the dominant marginal buyer of new dispatchable capacity in that state.

The pattern across today's brief is consistent. Alphabet is spending up to $205 billion this year to stay supply-unconstrained. OpenAI is building its own facilities to reduce dependence on third-party cloud. The question this raises for enterprise operators is structural: if the cost and availability of inference is now determined by who owns the power contracts and the silicon, not only who ships the best model, does your AI procurement strategy reflect that?