Elena Voss7.65 gigawatts. In the week ending August 14, Amazon backed an off-grid gas plant of that capacity for a Texas AI campus, despite its 2040 net-zero commitment. The plant could become the largest single source of emissions in the United States. Meanwhile, model input prices fell below $1 per million tokens. Digital intelligence got cheaper as the physical and political systems around it became harder to secure.
1. Capital arrived before energized sites
Nvidia joined Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR, and others on a $500 billion AI infrastructure funding package. Those firms are putting institutional balance sheets behind compute expansion, moving the risk beyond a handful of technology companies.
CoreWeave showed why the money arrived. Second-quarter revenue rose 112% year over year to $2.58 billion, while backlog reached $104 billion on 1.5 GW of contracted power. At roughly 40 times one quarter’s revenue, that backlog makes delivery the risk: CoreWeave must secure energized capacity before orders become cash.
The package can fund construction at industrial scale, but it also shifts AI-demand risk from model vendors to investors. Those investors can finance GPUs; they cannot manufacture energized sites. Capital travels globally, while generation remains local.
2. Cheap models rewarded distribution and efficiency
Gemini passed 1 billion monthly active users, becoming Google’s fastest-growing product and its fourteenth to reach that mark. Google then placed Gemini across the $899-and-up Pixel 11 line, making AI features central to otherwise modest hardware changes. Google can distribute a model through a billion-user app and its own devices without reacquiring each customer.
Google priced Gemini 3.7 Flash at $0.75 per million input tokens and $3.75 per million output tokens. DeepSeek priced V4-Pro at $0.435 and $0.87, respectively, charging about 23 cents for each dollar of Google’s output price. Sticker prices omit quality differences, but buyers still see useful intelligence selling for less than $1 per million input tokens.
Anthropic entered talks to acquire Decart for about $6 billion. Decart develops world models and GPU optimization technology intended to reduce training costs. The proposed deal would bring that optimization in-house as token prices fall.
Buyers gain leverage to demand portability, while vendors need default distribution or lower operating costs to preserve margins after the next model release.
3. Labs and Washington moved the permission line
Anthropic made auto mode the default in Claude Code for Pro, Max, and Team plans beginning August 14, saying it was good enough at identifying harmful actions. Users now supervise exceptions instead of approving each step.
Anthropic also said new Claude models will embed watermarks in generated text and C2PA metadata in files to comply with the EU AI Act. Those measures create evidence after the model acts, leaving prevention to the access and approval rules set beforehand.
OpenAI released GPT-5.6-Cyber, a more cyber-permissive version of GPT-5.6 Sol, to selected partners. Three days later, President Donald Trump authorized the US government to partner with private companies on cyberattacks abroad against criminal groups targeting Americans. OpenAI controlled access to a model while Washington opened operations to private firms.
As models execute longer sequences, companies and agencies rely more on choosing the operator than approving every step. Whoever grants access becomes part of the control system.
At first, 7.65 GW looked like an energy story. By week’s end, it measured the surrounding assets that remained scarce: power, distribution, provenance, and legal authority. Input token prices fell below $1 per million; Amazon still needed a Texas gas plant. That gap is what 7.65 gigawatts measured.