The Information Machine
The day in review

Monday 17 August 2026

4Moved
39New this week
93On the record

What moved

01
Concluded today

DeepSeek V4-Pro open-weight release

  • DeepSeek launched V4-Pro-0813 on August 13 across API, app, and web with three thinking tiers, OpenAI Responses API support, and Codex integration, publishing open weights under an MIT license on Hugging Face and raising peak-hour output token prices to $3.96 per million from a prior flat rate of $0.87.
  • Third-party benchmark comparisons placed it at 1607 on Code Arena WebDev AutoEval, second among open models on scores from a reward model rather than live human votes, with Kimi K3 costing $18 against DeepSeek's $1.305 for the same token volume.
The gist

DeepSeek V4-Pro is an open-weight model scoring near proprietary frontier models on several coding and agentic benchmarks at substantially lower API cost. The weights carry an MIT license, making the model self-hostable.

Continue this story →12 sources
02
Concluded today

OpenAI frontier enterprise token consumption data

  • Three enterprise case studies published in August 2026 document ChatGPT Work at Zapier for marketing automation, at Virgin Atlantic compressing competitive airline research from weeks to hours, and at OpenAI's finance team, where the CFO set a zero-day financial close as a goal.
  • These accompany OpenAI data showing frontier enterprises generated 8.3 times more output tokens per active user than typical firms as of mid-2026, up from 2.6 times in January, with non-developer Codex adoption rising 189 times from August 2025.
The gist

The gap between frontier and typical enterprise AI users has widened sharply in under a year, and usage is now spreading from engineering into legal, sales, marketing, and finance, indicating enterprises are moving AI from assistive tasks toward more delegated, agentic work.

Continue this story →7 sources
03
Concluded today

Columbia's US AI infrastructure spending research

  • A Columbia Business School research brief on August 16 warned the US could enter recession if data center investment falters, extending Van Nieuwerburgh's calculation that AI infrastructure will average 2.8% of GDP from 2025 to 2032, more than any prior US buildout.
  • Reuters reported AI issuance may push investment-grade totals above $2 trillion in 2026 for the first time, with investors questioning absorption capacity.
  • Regional grid data from ERCOT and EPRI corroborated BloombergNEF's 194-gigawatt US data center capacity forecast for 2035.
The gist

The 2.8% GDP share, if sustained, would make AI the largest single-sector capital commitment in US economic history, and Columbia's research implies the economy has grown dependent enough on the buildout that a slowdown carries recession risk. Debt markets are being pushed toward volumes that may cross historical issuance records, with investor absorption capacity already in question.

Continue this story →9 sources
04
Concluded today

Anthropic's data center lease with Riot Platforms

  • On August 17, Anthropic signed a 20-year, $9.1 billion lease with Riot Platforms for 191 megawatts of AI computing capacity in Rockdale, Texas.
  • Bloomberg identified Anthropic as the counterparty after Riot described the buyer only as a 'leading frontier AI lab'.
  • The contract runs through June 2048, with optional extensions that could raise total value to $16.1 billion, with 96 megawatts due by December 2027 and full capacity by June 2028, backed by $573 million in interim financing from Morgan Stanley.
The gist

The deal commits Anthropic to a substantial long-term infrastructure position, with a base value of $9.1 billion and a potential total of $16.1 billion if extensions are exercised.

Continue this story →5 sources

No news that day

← 16 August 202618 August 2026 →
The daily email

Want this in your inbox?

I send a short email each morning with the stories that moved. If you would rather just read here, that works too.

Subscribe free