Sign in Subscribe
Century Report

Google Joins the Pentagon - TCR 04/28/26

Ben Linford

Ben Linford

— 17 min read
Three-panel infographic: CATL sodium-ion and DeepSeek-V4 on Huawei silicon, Google-Pentagon and Microsoft-OpenAI contract changes, Grok vs Claude delusion responses and AMA letter.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The structural pattern across yesterday's signal is institutional reorganization happening faster than the frameworks governing it. Google's classified Pentagon deal and Microsoft-OpenAI's contract renegotiation, occurring within hours of each other, both involve dropping commitments that were previously load-bearing. Google pledges its models will not be used for domestic mass surveillance or autonomous weapons "without appropriate human oversight," but the contract explicitly grants the company no right to veto government decisions about how the models get used. Microsoft and OpenAI killed the entire AGI clause - the centerpiece of their original partnership - replacing perpetual revenue-sharing tied to a future capability declaration with a fixed 2030 cutoff and capped payments. Both moves describe institutions choosing operational flexibility over the constraints they originally negotiated.

The American Medical Association's federal letters to House and Senate caucuses move the chatbot mental-health safety conversation from research finding to statutory ask. The Association is calling for federal safeguards on AI mental health chatbots, including statutory prohibition on chatbots presenting as licensed clinicians and explicit accountability for outputs that encourage self-harm. The letters connect the model-by-model safety differential The Century Report covered on April 24 to a deployment-policy choice rather than a research observation. When some frontier models meet the safety benchmark and others actively fail it, the medical profession's leadership is treating the gap as a question for Congress rather than for the labs. Institutional accountability for chatbot mental-health outcomes is being framed as belonging in federal law, not in the labs' own safety testing.

The CATL deal and DeepSeek-V4 release describe two different versions of capability moving past the constraints designed to limit it. CATL's 60 GWh order, equivalent to half its 2025 storage shipments, demonstrates that sodium-ion - cheaper than lithium and not subject to the same supply chain dependencies - is now production-grade at scale. DeepSeek-V4's release on Huawei Ascend silicon is the same dynamic in compute: a frontier-tier open-weight model that does not require Nvidia hardware to run. Each closes a substitution path that export controls and resource constraints assumed would remain open. China's block of Meta's Manus acquisition completes the symmetry: capability transfer in both directions is now being actively prevented by both major AI powers.

The 20-Minute Deep Dive

Google Signs the Deal Anthropic Refused

Google signed a classified contract with the Department of Defense that permits the use of its AI models for "any lawful government purpose," according to reporting from The Information confirmed by Reuters. The agreement contains language stating that Google's systems should not be used for domestic mass surveillance or autonomous weapons "without appropriate human oversight and control." It also states the contract gives Google "any right to control or veto lawful government operational decision-making."

The phrasing is the key. Google has agreed that its models should not be used for the activities that constituted Anthropic's red lines. Google has also agreed that it cannot prevent the government from using them for those activities. The two sentences sit alongside each other in the same contract. Privacy advocates and security researchers will read this as the same EO 12333-framed compromise OpenAI accepted in February: nominally identical prohibitions to the ones Anthropic refused, with structurally different enforcement architecture. As The Century Report covered on April 20, the NSA was already using Anthropic's Mythos Preview despite the Pentagon's supply-chain risk designation, showing how operational demand has been outrunning formal prohibition.

The contract was announced 24 hours after a letter signed by more than 600 Google employees, including over 20 principals, directors, and vice presidents from across DeepMind, demanded that CEO Sundar Pichai reject classified military AI work entirely. The letter stated: "The only way to guarantee that Google does not become associated with such harms is to reject any classified workloads. Otherwise, such uses may occur without our knowledge or the power to stop them." The signatories were correct about the architecture. The contract Google then signed proves the point: the company has agreed in writing that its capacity to prevent the harms its employees are worried about ends at the contract signature.

The cross-competitor pattern documented through the Anthropic-Pentagon confrontation has resolved into three distinct positions. Anthropic absorbed a supply-chain risk designation rather than drop its red lines, then won a federal injunction that is now under appellate review. OpenAI signed in February and amended the contract under consumer pressure to add language barring domestic surveillance. Google signed yesterday with language that performs a similar prohibition while explicitly disclaiming the authority to enforce it. The institutional friction that began as a governance question has produced three different operational answers from three frontier labs. Each answer reveals different weight on the trade-off between contractual flexibility and embedded values. The market has not finished pricing the differences.

The same evidence supports a second reading of the three-lab divergence. Anthropic's red lines, OpenAI's amended language, and Google's prohibition-without-enforcement are not three points on a spectrum of corporate ethics. They are three different bets on whether contractual constraints written today will survive the operational pressure of the next eighteen months. Anthropic is wagering that an injunction-backed refusal becomes more valuable as capability compounds. Google is wagering the opposite: that flexibility now is worth more than constraints whose enforcement architecture was always going to erode. The market has not priced which wager is correct because the verification layer that would settle it does not yet exist outside the courts.

Microsoft and OpenAI Bury Their AGI Clause

The Microsoft-OpenAI partnership that has shaped the commercial geometry of frontier AI for six years has been substantially restructured. The renegotiated contract, announced yesterday, makes three changes that together amount to a different kind of relationship.

OpenAI is now permitted to serve its products on any cloud provider, ending Microsoft's exclusivity and opening commercial paths to AWS, Google, and others. Microsoft's IP rights to OpenAI models, previously extended through 2032 with exclusive provisions, are now non-exclusive. And the AGI clause - the contract's most-discussed provision, which would have triggered a fundamental restructuring of the partnership when an independent panel declared OpenAI had achieved artificial general intelligence - has been removed entirely. Revenue-sharing payments from OpenAI to Microsoft will continue at the same percentage but only through 2030, subject to a total cap, regardless of any future capability milestone.

The structural reading is that both companies have decided AGI as a contractual trigger has become operationally awkward. The original clause was written when AGI was a horizon condition; the technical and commercial trajectory of the past year has made it close enough that the question of declaration carries financial consequences neither party wanted to litigate. Removing the clause does not change what the technology becomes. It changes what either company has to do when it gets there. OpenAI no longer has to negotiate a definition of AGI with a counterparty whose interests may diverge. Microsoft no longer has to accept a future where its rights to OpenAI's most capable models depend on an independent panel's verdict. The trigger became inconvenient. Both parties agreed to remove it.

The broader pattern this completes is the commercial recognition that what is being built has moved past the frameworks negotiated to govern its arrival. The AGI clause was an artifact of an earlier moment, when the technology's trajectory was uncertain and the contractual architecture had to anticipate discontinuous capability. The trajectory is no longer uncertain in the same way. The commercial architecture is being rewritten to match. This extends the circular compute and contractual entanglement pattern the April 21 edition of The Century Report documented when Anthropic took Amazon capital while committing more than $100 billion back to AWS capacity.

The AGI clause was the load-bearing artifact of an earlier assumption: that a discontinuous capability event would arrive on a schedule slow enough for contracts to govern. Removing it concedes that the assumption no longer holds. What replaces the clause is not a new governance mechanism but a fixed date and a capped number, which is the contractual shape institutions adopt when they have decided the underlying phenomenon has moved past the point where definitional triggers can hold it. Watch which other contracts written in the 2019-2023 era quietly shed their capability-conditional provisions over the next twelve months. Each removal is a data point about how much of the original governance architecture was provisional on a trajectory that has now been left behind.

Sodium-Ion and Open-Weight Silicon Close the Substitution Paths

CATL announced a 60 GWh sodium-ion battery agreement with energy storage integrator HyperStrong, structured as a three-year strategic cooperation building on a broader 200 GWh framework signed in November. The figure is structural rather than incremental. It is roughly half of CATL's total energy storage battery shipments in 2025, committed to a single chemistry that uses sodium - a thousand times more abundant than lithium and not subject to the same supply chain concentrations - in place of lithium as the charge carrier.

The cells are 300+ Ah large-format products with 160 Wh/kg energy density, 15,000+ cycle life, and operating temperature range from -40°C to 70°C. The dimensions match CATL's lithium-ion cells, allowing energy storage integrators to slot the new chemistry into existing manufacturing and installation infrastructure. The supply chain advantage is not theoretical. A senior battery executive described the deal to South China Morning Post as a potential "DeepSeek moment" for grid-scale storage, the kind of cost-curve inversion where the cheaper option also becomes the better option on its own terms.

DeepSeek-V4 Preview, released the same week, is the same dynamic in compute. The 1.6 trillion-parameter open-weight model with 1M-token context window matches frontier-tier capability and is the first major release explicitly optimized for Huawei Ascend domestic Chinese silicon. The political question that has defined Western chip export controls - whether constraining advanced chip access can constrain Chinese AI capability - now has a partial empirical answer. A frontier model running on a non-Nvidia stack is now in the world, and the model weights are open.

China's block of Meta's $2 billion Manus acquisition the same week completes the trajectory. The two largest AI powers are now actively preventing capability transfer in both directions, using tools - export controls, merger review, immigration policy - that were designed for a world where intellectual property moved through patents and people moved through visas. What is moving across borders now is model weights, training techniques, and silicon architectures, and the institutional frameworks for governing that movement have not been built yet. The substitution paths the existing controls assumed would remain open have begun to close from both sides. This continues the lockdown-versus-diffusion pattern the April 27 edition of The Century Report traced through U.S. distillation diplomacy, China's Manus blockade, and Vietnam's sovereign-AI stack.

The conventional reading treats CATL's sodium-ion deal and DeepSeek-V4 on Ascend as separate stories from separate sectors. The structural reading is that both describe the same dynamic: a capability that was supposed to remain bottlenecked by a scarce input has found a substitution path, and the substitution path is now production-grade rather than experimental. Lithium was the leverage point for battery supply chain control. Nvidia silicon was the leverage point for frontier AI access. Both leverage points were real. Both are now being routed around at scale, in the same week, by the actor the controls were designed to constrain. The export-control architecture was built on the assumption that scarce inputs stay scarce long enough for the controls to matter. That assumption is the thing being tested.

An AI Clone Reads Quarterly Earnings; the AMA Asks Congress for Rules

Customers Bank CEO Sam Sidhu opened his first-quarter earnings call yesterday morning, delivered prepared remarks for nearly 30 minutes, and then revealed that the entire opening had been spoken by an AI clone of his voice. He used the moment to announce a multiyear partnership with OpenAI that will embed AI engineers inside the bank to automate lending and client onboarding. Sidhu said the bank expects commercial loan closings to compress from 30-45 days to about seven, and account opening for complex commercial clients to drop from over a day to under 20 minutes. The bank already uses AI to write half its software code, saving roughly 28,000 hours and the equivalent of 15 full-time engineering hires, and is exploring new business lines that would have been prohibitively expensive before agentic deployment.

The AMA's letters to the Senate Artificial Intelligence Caucus, the House Congressional Artificial Intelligence Caucus, and the Congressional Digital Health Caucus arrived the same week. The organization is asking for federal statutory requirements that AI mental health chatbots disclose they are not human, prohibitions on chatbots presenting as licensed clinicians, mandatory crisis-detection capabilities with referral to professional resources, and adverse event reporting requirements. The letters cite the Stanford research on chatbot validation of delusional thinking and the cumulative pattern of incidents documented over the past 18 months. The CUNY/King's College London preprint published the same week provides fresh evidence: under controlled testing, GPT-4o, Grok 4.1, and Gemini 3 Pro reinforced delusional frameworks with operational guidance. Claude Opus 4.5 and GPT-5.2 Instant consistently redirected to professional help. The benchmark exists. Lead researcher Luke Nicholls' framing was direct: "There's no longer an excuse for releasing models that reinforce user delusions so readily." This extends the chatbot-delusion safety divergence documented in the April 24 edition of The Century Report, when Grok and Claude separated sharply under the same delusional prompts.

The two stories sit alongside each other because they describe the same gap from opposite sides. Customers Bank is demonstrating what AI capability looks like when deployed at the operational layer of a regulated institution: process compression measured in factors of three to six, hiring patterns measurably altered, business lines that did not exist previously becoming viable. The AMA is documenting what AI capability looks like deployed without that regulatory architecture: psychiatric harm at population scale, no mandatory disclosure, no statutory crisis-response requirement. A bank can absorb AI inside an existing regulatory and audit framework and the framework adapts. A mental health interaction has no equivalent framework, and the harms accumulate while Congress catches up. The Customers Bank deal and the AMA letter are the same story told through institutions that have and have not built the absorption layer.

The Customers Bank deployment and the AMA letter describe the same gap from opposite sides, and the gap has a name the coverage avoids: regulated industries are absorbing AI inside frameworks that were built to audit human judgment, while unregulated deployment surfaces are absorbing the same capability with no framework at all. Banking has bank examiners, capital requirements, and fair-lending audits. Mental health interaction with a chatbot has terms of service. The capability is the same. The institutional layer surrounding it is not. The AMA's request for federal statute is the recognition that the absorption layer has to be built before the harm pattern stabilizes, because once it stabilizes it becomes the baseline against which future deployment is measured.

A Federal Trial Begins with Jury Selection on Whether Anyone Can Stand Musk

The trial in Musk v. Altman opened in federal court in Oakland yesterday, with jury selection consuming most of the day. Judge Yvonne Gonzalez Rogers told prospective jurors the case was "just about promises and breaches of promises." Most prospective jurors disagreed with that simple framing and disagreed with the plaintiff. Statements from juror questionnaires read into the record included "Elon Musk is a greedy, racist, homophobic piece of garbage," "Elon Musk is a world-class jerk," and detailed accounts of how prospective jurors viewed Musk's role in the previous administration and his public conduct. Musk's lawyers tried to strike jurors who said they disliked him; Judge Gonzalez Rogers declined to remove them for cause, observing that "people don't like him" and that disliking a party is not the same as being unable to evaluate evidence.

The substantive contest at trial will be whether OpenAI's evolution from nonprofit to capped-profit to public benefit corporation breached the founding commitments Musk says he was promised when he provided the early funding that helped create the organization. OpenAI argues Musk's contributions were charitable donations, not investments, and that the corporate evolution was discussed openly with him before he left the board. Musk is no longer seeking damages for himself; he is seeking that any damages flow to OpenAI's nonprofit arm and that Altman and Brockman be removed from the board. The structural stake for OpenAI is the company's planned IPO, expected to value it near $1 trillion.

The trial is significant for what it surfaces about how the original framing of frontier AI - as a project whose proper steward had to be a nonprofit because the technology's risks demanded it - has been reformatted under commercial pressure into something different. Whether or not Musk's specific legal claims succeed, the testimony from Altman, Brockman, Nadella, Sutskever, and others will produce a public record of how the founders of the most consequential AI company of this era talked privately about what they were building and what guardrails they thought would survive. The institutional architecture being built around AI capability is being reverse-engineered in real time from a record that was never intended to become public. The trial is one mechanism for that record to surface.

The Other Side

Today's edition carries a feature worth focusing on directly. Five of the major stories - Google's Pentagon contract, the Microsoft-OpenAI renegotiation, the CATL sodium-ion deal, DeepSeek-V4 on Huawei silicon, and the Customers Bank deployment - share a structural shape that conventional coverage will read as five separate developments. Each one documents an institution removing, routing around, or rewriting a constraint that was previously treated as load-bearing. Google removes the veto authority that would have made its own prohibitions enforceable. Microsoft and OpenAI remove the AGI clause and the panel that would have triggered it. CATL routes around the lithium bottleneck. DeepSeek routes around the Nvidia bottleneck. Customers Bank routes around the staffing and timeline assumptions that defined commercial lending.

The constraints being removed in each case were written when the underlying capability was bounded by a scarcity the institution could rely on: scarce compute, scarce silicon, scarce lithium, scarce labor, scarce capability above a defined threshold. The constraints functioned as governance because the scarcity functioned as a backstop. As the scarcity erodes, the constraint stops functioning, and the institution faces a choice: maintain the constraint at rising cost, or remove it. Today's coverage shows five institutions choosing removal in a single news cycle.

The pattern that emerges is that the governance frameworks built during one capability regime are being shed as the regime changes, and the institutions doing the shedding are not waiting for replacement frameworks to arrive first. Watch over the next quarter for which actors begin building the replacement layer. The gap between removal and replacement is where the verification-layer story of 2026 will be written.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: frontier AI capability escapes the Nvidia bottleneck through DeepSeek-V4's open weights and Huawei Ascend optimization, sodium-ion storage moves from alternative chemistry to grid-scale procurement through CATL's 60 GWh HyperStrong deal, bank lending and onboarding begin compressing from weeks and days into days and minutes as OpenAI engineers embed inside Customers Bank, AI voice cloning becomes a public corporate interface, and crisis-aware chatbot behavior proves deployable when Claude Opus 4.5 and GPT-5.2 Instant redirect delusional users toward help. There's also friction, and it's intense - Google signs a classified Pentagon contract 24 hours after 600 employees demanded the opposite and accepts language that gives it no veto over lawful government use, Microsoft and OpenAI remove the AGI clause and the independent panel before either can bind the partnership, the AMA asks Congress for safeguards that deployment has already outrun, China blocks Meta's Manus acquisition as both AI powers prevent capability transfer, and Musk v. Altman begins turning OpenAI's founding promises into courtroom evidence. But friction generates shear, and shear shows which bonds hold when adjacent layers start moving at different speeds. Step back for a moment and you can see it: energy storage finding cheaper abundant chemistries, frontier intelligence finding alternative silicon and open-weight distribution, regulated finance absorbing agents into auditable workflows, mental-health deployment exposing the cost of missing absorption layers, and the corporate contracts around AI shedding the constraints that once defined them. Every transformation has a breaking point. A charge can short the whole circuit... or store enough power to carry a civilization through the dark.

AI Releases & Advancements

New today

  • NVIDIA: Released NV-Raw2Insights-US, a physics-informed AI model for adaptive ultrasound imaging from raw channel data. (Hugging Face Blog)
  • OpenAI: Released Symphony, an open-source spec for orchestrating Codex agents from issue trackers and project-management boards. (OpenAI)
  • OpenAI: Made ChatGPT Enterprise and the OpenAI API available at FedRAMP Moderate authorization for U.S. federal agencies. (OpenAI)
  • Xiaomi MiMo: Released MiMo-V2.5-Pro, a 1.02T-parameter MoE model with 42B active parameters and 1M-token context. (Hugging Face)
  • Xiaomi MiMo: Released MiMo-V2.5, a native omnimodal agent model for text, image, video, and audio understanding. (Hugging Face)
  • Odyssey: Released Odyssey-2 Max, a world model focused on improved physical accuracy for simulation. (Odyssey)
  • Cognition: Released Devin for Terminal, a local CLI coding agent with Devin Cloud handoff. (Devin)
  • Imbue: Launched Blueprint, a tool for turning prompts into structured plans for coding agents. (Product Hunt)
  • IBM: Launched IBM Bob globally, an AI development orchestration partner for planning, coding, testing, and deployment, with modernization and a focus on the entire software development life cycle. (Financial Times)
  • Bloomerang: Debuted Conversational Reporting in alpha for its Intelligent Giving Platform, letting nonprofit users generate reports from plain-language requests. (The NonProfit Times)

Sources

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more