The Information Machine
The edition

Thursday November 12, 2026

This is one edition of my daily email.

One free email each morning, around 8 AM US Eastern.

In this edition

  1. New today
  2. NewGoogle DeepMind's AlphaGenome 2 releaseDeepMind Releases AlphaGenome 2; Original Model Applied to Alzheimer's Research
  3. NewAmazon's autonomous Ring Rover and Drone CamAmazon Launches AI-Powered Ring Rover and Indoor Security Drone
  4. NewOpenAI ChatGPT Business Profiles betaOpenAI Launches ChatGPT Business Profiles and Business Agent Beta
  5. NewOpenAI and Upwork freelance AI researchOpenAI and Upwork Joint Study Finds AI Projects Pay More, Span More Tasks
  6. Updates
  7. Day 24Anthropic IPO and compute commitmentsAnthropic Sets $70, $78 IPO Range, Roadshow Discloses Compute Multiplier Strategy
  8. Day 3Hyperscaler Q3 earnings and AI supply crunchAI Infrastructure Capex Projected to Top $2 Trillion by 2028
  9. Day 3Jensen Huang's alleged Chinese intelligence targetingDOJ Opens Criminal Probe of Megaspeed as Agencies Split on Blacklist
  10. Day 2Gemini 4 Carbon releaseGemini 4 Carbon Opens to All Paying Developers and AI Ultra Subscribers
  11. Day 2Machine-code formal verification of Signal's protocolDeepMind AlphaProof Code Formally Verifies Signal Cryptography at Machine-Code Level
  12. Day 2Anthropic's Claude workplace productivity toolsAnthropic Launches Workplace Suite Including Live Meeting AI
  13. Day 2US AI data center power requirementsMicrosoft Pursues 20-Year Chevron Deal for 2.67GW Data Center Power
  14. Day 1Hyperscaler data center commitments and credit stressBig Tech Data Center Commitments Top $3T as Credit Markets Show Strain
  15. Day 2Nvidia H200 access and shell company approvalsNYT: Shell Companies With Sanctioned Ties Among Commerce-Approved H200 Buyers

New today

01
New

Google DeepMind's AlphaGenome 2 release

  • Google DeepMind released AlphaGenome 2 on November 11, 2026, a conditional genomic model that takes both a DNA sequence and a target cell type as input and identifies causal variants in 30% of cases versus 16% for the prior model on the TenK10K immune cell dataset.
  • As a demonstration, DeepMind reported it reproduced a published finding that a variant drives inflammatory bowel disease by overactivating the ETS2 gene specifically in gut macrophages.
  • Separately, researchers have applied the original AlphaGenome to Alzheimer's disease research: one study nominated candidate causal genes in a GWAS of argyrophilic grain disease, and another found an AD risk variant may alter transcription and splicing of the Fc receptor gene FCER1G.
  • API outputs cannot be used to train other machine learning models, and commercial access is available via Google Cloud.
The gist

The model addresses a core challenge in human genetics: explaining how a single-letter DNA change can affect disease in one cell type while sparing others. Independent application of the original AlphaGenome to Alzheimer's research shows the framework is being adopted for mechanistic analysis of GWAS variants in neurodegenerative disease.

02
New

Amazon's autonomous Ring Rover and Drone Cam

  • Amazon announced two autonomous AI security devices: the Ring Drone Cam, a $249.99 indoor drone that deploys when alarms trigger and ships January 27, 2027 pending FCC authorization, and the Ring Rover Cam, a $999.99 rolling patrol robot for businesses that ships December 16, 2026.
  • Both use Ring's AI Pilot system, which combines vision models, on-device sensors, and cloud processing to map spaces, label rooms, and suggest patrol points such as doors and windows.
  • Amazon describes the Rover as "a guard who never clocks out" and acknowledges the Drone Cam's propellers are not quiet.
The gist

The devices bring autonomous AI-driven physical patrol to consumer and business home security, using continuous mapping and alarm-triggered deployment. The products are expected to intensify privacy concerns previously raised around Amazon's earlier attempt at a flying security camera.

03
New

OpenAI ChatGPT Business Profiles beta

  • OpenAI announced on November 10, 2026, a U.S. beta letting companies claim a presence in ChatGPT, manage their information, and view aggregated engagement metrics such as mention counts and website clicks, with no access to private user chats.
  • Verified businesses can also deploy a labeled Business Agent that answers customer questions using only company-supplied content, such as product feeds.
  • The initial rollout targets U.S. e-commerce and online-service companies.
The gist

The feature gives businesses a direct channel to manage how they appear in ChatGPT conversations and respond to shopper queries. The Business Agent operates exclusively on company-supplied information, which bears on how businesses control their representation in AI-mediated interactions.

04
New

OpenAI and Upwork freelance AI research

  • OpenAI and Upwork published joint research finding that AI-related freelance projects carry larger fixed budgets, bill more hours, combine rarer skill pairings, and cover more distinct tasks than comparable non-AI projects.
  • The research also identifies shifting demand for freelance work overall and new skill combinations emerging among freelancers as they adapt to AI-related work.
The gist

The research provides direct labor-market data on how AI adoption is reshaping freelance work patterns, including pay, hours, and skill composition. It shows AI-related projects differ structurally from non-AI ones, not just in topic.

Get the daily email

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

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy

Updates

05
Day 24

Anthropic IPO and compute commitments

  • The IPO roadshow video, released November 11, disclosed that an internal system called Mythos enabled Claude to port itself to AMD GPUs in three days, writing nearly all required kernels with almost no human intervention, and introduced 'total effective compute' as physical capacity multiplied by compute multipliers Anthropic believes Claude will eventually discover autonomously, accelerating toward recursive self-improvement.
  • Alongside the confirmed $70 to $78 price range, Anthropic announced a $26 billion private placement from Nvidia, Google, Microsoft, and four other suppliers.
The gist

An AI lab going public with over $1 trillion in compute commitments, while disclosing it may imminently cross its own definition of dramatically accelerated AI progress, puts both financial and safety questions in front of public markets simultaneously. The roadshow's disclosure that Claude is already autonomously accelerating infrastructure work, and that Anthropic expects this to compound, makes the RSP warning concrete rather than abstract.

06
Day 3

Hyperscaler Q3 earnings and AI supply crunch

  • CoreWeave reported Q3 2026 revenue of $3.7 billion, up 168% year-over-year with $9 billion in customer prepayments and margins ahead of plan.
  • Goldman Sachs estimated hyperscalers need roughly $300 billion annually to break even on capital expenditures, against Q2 2026 cloud growth at 48% with backlogs above $1.5 trillion; JP Morgan projected related capex exceeding $2 trillion by 2028.
  • SoftBank posted a quarterly loss of roughly 1.12 trillion yen driven partly by a $3.7 billion writedown on nonrefundable AI power payments.
The gist

The scale of committed capital, with backlogs exceeding $1.5 trillion and multiyear supply contracts with deposits covering 2027 and 2028 capacity, marks a structural departure from prior tech investment cycles. Supply constraints across HBM, EUV tools, chip packaging, and specialty memory mean the binding constraint is shifting from funding to physical manufacturing capacity.

07
Day 3

Jensen Huang's alleged Chinese intelligence targeting

  • CSET published an analysis arguing espionage allegations against Nvidia CEO Jensen Huang remain unproven and no evidence of technology transfer has emerged; researcher Sam Bresnick described the case as a sex scandal, an espionage issue, and Nvidia selling chips to Megaspeed, a firm investigators believe may not actually exist.
  • The inter-agency dispute over Megaspeed's export blacklisting sharpened when Treasury Secretary Bessent confronted Commerce Secretary Lutnick, with Axios reporting State and other agencies had agreed to the listing months earlier.
The gist

A criminal investigation into a company that bought over $4 billion in advanced Nvidia chips, combined with an inter-agency dispute over whether to blacklist it, puts Nvidia's export compliance and its CEO's conduct at the center of active federal enforcement. The board probe and DOJ investigation are running in parallel with unresolved questions about whether restricted chips reached China.

08
Day 2

Gemini 4 Carbon release

  • The Neuron Daily added a caveat that Carbon's top finishes on CursorBench, WANDR, and AutomationBench are on Google's own or partner-run tests, qualifying the significance of those first-place rankings from Carbon's November 10 general availability.
  • Carbon supports low, medium, and high thinking effort levels, defaulting to medium, with thinking and output tokens sharing the configured maximum.
  • The llm-gemini plugin was updated to version 0.36 to support Carbon.
The gist

Carbon's general availability makes Google's current frontier model accessible to developers and enterprises at introductory pricing that doubles after November 30. Its benchmark profile shows strength in cybersecurity and overall intelligence rankings but trailing positions in coding agent tasks, so the model's fit depends on the workload.

09
Day 2

Machine-code formal verification of Signal's protocol

  • Ars Technica reported that the cryptography AlphaProof Code verified covers not only Signal's libsignal but also Google Messages and WhatsApp, widening the scope of DeepMind's November 10 announcement.
  • Ars Technica characterized the work as a proof-of-concept and noted that practical scalability of AI-assisted formal verification remains uncertain.
The gist

Formal verification at the machine-code level closes the gap between protocol design and implementation, where real-world vulnerabilities typically hide. DeepMind cited medical device firmware and power grid software as candidate domains for the same approach.

10
Day 2

Anthropic's Claude workplace productivity tools

  • Claude Meetings, an additional tool in the November 10 workplace suite, joins Google Meet calls as a visible participant, drafts a plan during the call, and saves notes and action items afterward.
  • An admin must approve before it can join calls that include outside guests, and it leaves if one appears.
  • Coverage also confirmed that all tools in the suite require a paid plan and are in Team and Enterprise beta.
The gist

Anthropic is moving into workplace productivity software with tools that generate spreadsheets, dashboards, and meeting notes from plain-language requests rather than requiring users to build them manually. The launch sent Microsoft shares down more than 3.5%.

11
Day 2

US AI data center power requirements

  • Bloomberg reported that Microsoft is pursuing a 20-year power purchase agreement with Chevron for a 2.67GW gas plant to supply its data centers directly, committing that neither taxpayers nor other utility customers will cover the bill.
  • The arrangement fits the behind-the-meter self-generation model the Trump administration has pushed, but Texas Governor Greg Abbott's data-center permit moratorium has complicated it, and Abbott's Democratic successor has signaled a potentially tougher regulatory stance on such deals.
The gist

Federal policy, proposed legislation, and grid rules are converging to push large AI data center operators toward self-generation rather than dependence on the public utility grid. Microsoft's Chevron deal shows a major operator moving toward behind-the-meter generation at scale, while state-level permitting friction in Texas illustrates that the path remains contested.

12
Concluded today

Hyperscaler data center commitments and credit stress

  • Earnings guidance placed the Big-5 hyperscalers on track for $780 to $800 billion in combined 2026 capital expenditure, roughly three times 2024 levels: Amazon leads at approximately $220 billion, reaffirmed October 29, with Alphabet guiding $195 to $205 billion.
  • McKinsey projects $6.7 trillion in global data center capex by 2030; Dell'Oro forecasts $1.7 trillion cumulatively by that year.
  • Oracle's 5-year credit default swap spread has more than tripled since September 2025, and AI-related bonds now constitute 18% of the investment-grade corporate bond index, its largest single sector.
The gist

The pace of commitments has made AI-related bonds the largest single sector in the investment-grade corporate bond index, and Oracle's debt-funded approach has drawn visible credit market scrutiny. Access to power-ready, permitted land has become a material constraint on how fast the buildout can proceed.

13
Day 2

Nvidia H200 access and shell company approvals

  • The New York Times investigation identified Baidu, BYD, and Tesla Shanghai among roughly two dozen Chinese companies approved by the Commerce Department to buy Nvidia H200 chips, each licensed for tens of thousands of units though none have yet met those volumes.
  • Corporate records show the vetting problems extend beyond New Matrix and Maginfra, the two apparent shell companies with sanctioned ties, to other approved firms connected to U.S.-blacklisted entities or military contractors in China.
The gist

The Commerce Department's approval process cleared firms with apparent ties to entities it had separately blacklisted, raising questions about whether export screening has systemic gaps. Approved licenses covering tens of thousands of chips each remain open, and the chips are central to AI data center infrastructure.

What is moving now · All editions · All stories