AI is crystallizing around a few choke points: agents that can move money on their own, scarce and politically exposed compute, and vertically integrated stacks from labs like OpenAI. Governments are starting to wall off sensitive data—sometimes handing it to players like Palantir—while workers and the public push back even as code and operations quietly get automated.
The real risk now is less about raw model capability and more about who ends up owning the rails, the power, and the data.
Key Events
/OpenAI created a majority‑owned Deployment Company with ~$4B from 19 partners and ~150 engineers to help businesses deploy its models.
/OpenAI will wind down its fine‑tuning API and require customers to migrate off the service by January 2027.
/Palantir was granted 'unlimited access' to identifiable NHS England patient data, prompting MPs to label the deal 'dangerous.'
/Texas sued Netflix for allegedly spying on children, fostering addiction, and misleading users about data collection and advertising.
/Florida and Oregon moved against AI data centers, allowing citizen challenges and forcing facilities to pay full local grid‑expansion costs.
Report
Capital is herding into two AI chokepoints: agents that can transact on their own, and the costly, politically exposed infrastructure that keeps them running.
Everything else this month—OpenAI’s new deployment arm, Palantir’s NHS deal, EU moves against hyperscalers, and Samsung’s strike risk—sits somewhere on that fault line.
agents as the new transaction layer
AWS launched Bedrock AgentCore Payments, letting AI agents initiate payments through partners like Coinbase and Stripe and even hold wallets for autonomous spend.
In parallel, OpenAI is spinning up a majority‑owned Deployment Company with about 150 engineers and $4B from 19 partners to embed its models directly into business workflows.
GitLab is explicitly cutting jobs to reorient around what it calls the 'agentic era,' signaling that mainstream dev tooling vendors now see agents as central to their growth.
Underneath that, coding agents are being benchmarked across 150 realistic tasks, with tools like Claude Code and cheaper rivals such as Kimi pushing performance up and prices down.
the compute and data‑center crunch
An AI data center quietly used 29 million gallons of water over 15 months before being called out, putting a hard number on AI’s resource draw.
Florida now lets citizens challenge AI data centers, while Oregon forces facilities to pay the full cost of local grid expansion and related upgrades.
To stand these sites up quickly, operators are leaning on older, dirtier energy sources, even as Mitsubishi Heavy Industries retools gas‑turbine production specifically for AI data‑center demand.
Chip supply remains shaky, with Samsung warning an 18‑day fab strike could cost $20B and TSMC declining ASML’s latest tool because the economics do not clear.
openai and the frontier stack land grab
OpenAI’s Deployment Company shifts it from API vendor to implementation partner, backed by outside capital but majority‑owned, tightening its grip on enterprise deployments.
The same month, it announced Daybreak for continuous cyber defense using its models and decided to sunset its fine‑tuning API, pushing customers off by January 2027.
OpenAI is also giving the EU access to a new cyber model even as rival Anthropic withholds its equivalent, underscoring divergent regulatory strategies between labs.
Anthropic, meanwhile, has its Claude API on AWS with same‑day parity, faces warnings that 100% of its secondary trades are illegal, and sees a low‑cost competitor Kimi offering a 'Claude‑like' experience at about one‑sixth the price.
sovereign data, palantir, and the hyperscaler wall
Palantir has been granted what officials describe as 'unlimited access' to identifiable NHS England patient data, including handling by its contractors, triggering MPs to brand the arrangement 'dangerous.' Commenters across Reddit and Hacker News are calling the firm everything from untrustworthy to a 'terrorist organization' and demanding a 'Palantir Privacy Act' to rein it in.
Despite assurances about air‑gapped infrastructure and following NHS instructions, critics question whether the health service even has authority to grant this access without explicit patient consent.
In parallel, the EU is moving to block Microsoft, Amazon, and Google from holding sensitive government data, effectively carving out a separate track for 'sovereign' cloud even as European investments into Palantir increase.
ai‑written code, productivity theater, and cultural blowback
Airbnb now says AI writes about 60% of its new code, a stat widely repeated by both tech press and developer communities. Developers report 'vibe coding' fatigue and harder debugging as they inherit complex AI‑generated code they do not fully understand, even as coding agents expand into review, QA, and infrastructure monitoring.
Inside big tech, Meta employees describe AI‑driven changes as making them miserable and anxious, while a BCG study coins 'AI brain fry' for decision overload from juggling multiple agents in workflows.
Public sentiment is fraying, with students booing commencement speeches about AI as the 'next industrial revolution' and younger generations saying the current AI hype ignores issues like housing, climate, and healthcare.
What This Means
AI is converging into a few chokepoints—agentic transaction rails, constrained compute, and politically gated data—and the biggest financial outcomes will come from how those bottlenecks get priced and who controls them. The tension between hyperscale, sovereign, and local models is no longer theoretical; it is already rewriting contracts, labor, and infrastructure economics in real time.
On Watch
/Miami startup Subquadratic’s claim of a 1,000× efficiency gain with its SubQ model is drawing attention but also public demands from researchers for independent validation.
/MIT’s FINGERS‑7B multi‑omics model for Alzheimer’s prevention, trained on data from 30,000 people, is an early test of how far high‑stakes health AI can go in mainstream clinical settings.
/Google’s leaked Omni model, the planned Flash 3.2 release, and Gemini‑based video editing together hint at a near‑term step change in multimodal capability ahead of I/O.
Interesting
/The commoditization of the app layer in technology suggests a shift towards infrastructure, as competition among models approaches saturation.
/Anthropic's Claude was implicated in an attempted compromise of a Mexican water utility, raising security concerns.
/Local open-weight AI on laptops is reportedly improving at a rate exceeding Moore's Law, indicating rapid advancements in model performance.
/Governments' recognition of frontier AI models as critical infrastructure marks a pivotal change in cybersecurity strategy.
/Cowboy Space raised $275 million to build rockets aimed at supporting space data centers, indicating a new frontier in data storage solutions.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/OpenAI created a majority‑owned Deployment Company with ~$4B from 19 partners and ~150 engineers to help businesses deploy its models.
/OpenAI will wind down its fine‑tuning API and require customers to migrate off the service by January 2027.
/Palantir was granted 'unlimited access' to identifiable NHS England patient data, prompting MPs to label the deal 'dangerous.'
/Texas sued Netflix for allegedly spying on children, fostering addiction, and misleading users about data collection and advertising.
/Florida and Oregon moved against AI data centers, allowing citizen challenges and forcing facilities to pay full local grid‑expansion costs.
On Watch
/Miami startup Subquadratic’s claim of a 1,000× efficiency gain with its SubQ model is drawing attention but also public demands from researchers for independent validation.
/MIT’s FINGERS‑7B multi‑omics model for Alzheimer’s prevention, trained on data from 30,000 people, is an early test of how far high‑stakes health AI can go in mainstream clinical settings.
/Google’s leaked Omni model, the planned Flash 3.2 release, and Gemini‑based video editing together hint at a near‑term step change in multimodal capability ahead of I/O.
Interesting
/The commoditization of the app layer in technology suggests a shift towards infrastructure, as competition among models approaches saturation.
/Anthropic's Claude was implicated in an attempted compromise of a Mexican water utility, raising security concerns.
/Local open-weight AI on laptops is reportedly improving at a rate exceeding Moore's Law, indicating rapid advancements in model performance.
/Governments' recognition of frontier AI models as critical infrastructure marks a pivotal change in cybersecurity strategy.
/Cowboy Space raised $275 million to build rockets aimed at supporting space data centers, indicating a new frontier in data storage solutions.