The Government Bought Frontier AI for a Dollar
The federal government put the most capable commercial AI models inside its agencies last year at about a dollar a seat. That price is the tell. In federal health the value moved to everything the dollar did not buy: the platforms the models ride on, the accreditation that lets them run, the governance that keeps them safe, and the encryption that has to outlive the patient.
Friends,
In August 2025 the federal government acquired the most advanced commercial AI models on the market for roughly the price of a vending-machine snack. OpenAI's ChatGPT Enterprise came in at a dollar per agency. A week later Anthropic's Claude matched it. Google priced Gemini for Government at forty-seven cents, and by late September xAI's Grok undercut them all at forty-two.
Those prices ran through GSA's OneGov strategy, which treats the whole government as a single buyer and pushes the discount through the federal schedule. By the end of 2025 GSA had executed nineteen of these agreements at discounts reaching ninety percent. By May 2026 it counted twenty companies, more than a hundred and twenty orders, and roughly 3.4 million users, against a reported first-year saving above a billion dollars. The frontier model, the thing two governments are racing each other to build, entered the United States government priced like a rounding error.
That dollar sign is the most important number in federal health IT this year, and almost none of the coverage read it correctly.
The race became policy
The dollar did not come out of nowhere. It came out of a deliberate turn in Washington that recast the technology contest with China from a private performance race into national-security and industrial policy.
The chain is on the record. In January 2025 EO 14179 ordered the removal of barriers to American AI leadership. That July the administration published its AI Action Plan, "Winning the Race," which named healthcare as a slow-adopting sector to be accelerated through NIST standards, testbeds, and FDA regulatory sandboxes. In June 2026 EO 14409 stood up a voluntary pre-release review for frontier models and an AI Cybersecurity Clearinghouse, and a national-security memorandum, NSPM-11, folded AI into the defense enterprise as a readiness capability. The money moved with the language. The FY2027 request scales AI-related defense spending toward $58.5 billion.
One fact inside that chain does most of the work for anyone in federal health, and it is a fact about absence. Not one AI executive action from 2025 or 2026 names HHS, CMS, or the VA, and none carries a defense-health provision. The policy that reorganized the government around AI has no health-specific instruction in it. The roll-down to health is happening through agencies implementing government-wide mandates on their own, not through a directive written for the mission.
You can watch it happen. CMS stood up a new Office of Health Technology and Products in June 2026 to centralize AI strategy, interoperability, and digital-product development, with no budget or staffing figures published alongside it. That is an agency building its own AI machinery because the top-line policy handed it a mandate and left the how to the building. Every health agency is doing some version of the same thing, at its own pace, on its own dime.
That is the first clue to where the value went. When the policy hands you a mandate and no blueprint, the blueprint becomes the product. The model was priced at a dollar in Washington. The instructions for running it safely in a hospital were not priced at all, because no one at the top wrote them.
A dollar buys the model and nothing else
That dollar is an introductory promotional rate, one price per agency, and it runs out. Once the window closes, the standard rate for Claude for Government is sixty dollars a seat per month. The dollar buys the model through the door. It buys none of the work that turns a model into something a federal system can actually run: the integration, the secure hosting, the accreditation to a federal authorization level, the training, the model consumption at scale.
The model itself confers no advantage anymore, because the government bought all of them. GSA put four frontier vendors on the schedule inside two months. The Department of War, on May 1, 2026, cleared eight firms to run their AI on classified networks. When four vendors sit behind the same schedule and eight sit behind the same classified clearance, no single foundation model is a moat. The capability everyone spent five years arguing about is now a commodity at the point of federal entry.
Which raises the question a capture lead should already be asking. If the model is a dollar and confers no edge, where did the money go.
It went down-stack, into the layers the dollar did not touch.
The tool is a line item; the platform is the program
Start with the clearest case in the record. In February 2026 the Department of Veterans Affairs put an ambient clinical scribe into production, the tool that listens to a patient visit and drafts the note so the clinician is not typing at midnight. The award went to Rise8, a service-disabled veteran-owned small business, teamed with Thoughtworks, on a sole-source contract through the VA Technology Acquisition Center. The commercial engines underneath are Abridge and Knowtex. The total value is $4,945,717.92.
Now set that number against the platform it depends on. The VA scribe rides on the Electronic Health Record Modernization program, the Oracle Health rollout that VA now internally estimates at $37.2 billion, of which roughly $12.7 billion is already spent, with Accenture Federal Services as the integrator and deployments resuming this fiscal year. The scribe is a five-million-dollar tool sitting on a thirty-seven-billion-dollar platform. It is a rounding error on the thing that makes it possible.
The defense side runs the same shape. DHA fielded its own ambient documentation tool, the Clinical AI Agent, through a six-week limited fielding at Madigan, Camp Lejeune, Portsmouth, and Wilford Hall in late 2025, then began a phased enterprise rollout in February 2026. At Camp Lejeune providers cut about 1.2 hours a day of charting and a quarter of their after-hours work. The vendor and the dollar value are not public, because the tool was competed as an Other Transaction agreement under a commercial solutions opening rather than a traditional contract. What is public is the platform it rides on: MHS GENESIS, the military health record whose original integration award to Leidos ran $4.3 billion in 2015 and rose to $5.5 billion after the 2018 modification.
The pattern holds across both departments. The AI tool is a small line item, sometimes undisclosed. The platform underneath it is a program measured in billions. Anyone selling the tool is selling the cheap part, and the durable money sits in the platform, the integration, and the sustainment that keeps both running.
The largest window on the board just opened at DHA
The clearest proof that the value sits below the model is the market DHA is opening right now, and it has nothing to do with AI at all.
DHA is ending Leidos's role as the lead systems integrator for MHS GENESIS, a job Leidos has held since 2015. In its place the agency plans to award five sole-source, firm-fixed-price, outcome-based contracts directly to the platform providers: Oracle Health for the core record, Philips for tele-critical care and remote monitoring, Amwell for virtual care, Henry Schein for dental, and Solventum for clinical documentation and coding. Integration comes in-house. And everything that used to sit behind the Leidos prime, the program-management support, product management, deployment, training, test, data analytics, cyber, and help desk, gets competed full and open.
The agency's own language for why is worth quoting plainly. It is unwinding the arrangement to eliminate the legacy integrator's pass-through costs. Translated, DHA decided it was paying a prime to mark up the work of subcontractors it could contract with directly, and it stopped.
The transition runs on a published clock. Philips and Amwell move by the end of July 2026. Oracle by November 2026. Henry Schein and Solventum by July 2027. The platform seats go sole-source and are effectively spoken for. The services layers are the prize. Those were locked under the Leidos prime for a decade, and they come onto the open market between now and 2027. On the VA side the analogous production work went to a service-disabled veteran-owned small business, which tells you the services layer is set-aside-friendly. This is the largest concentration of newly competable defense-health work in years, and it is a systems-integration and sustainment market, not a model market.
Governance is the constraint the OIG already documented
The same layer that holds the money holds the risk, and the risk now has a paper trail.
In January 2026 the VA Office of Inspector General issued a preliminary advisory memorandum on VHA's use of generative AI. It found that clinical staff were using generative tools, including the department's own VA GPT and Microsoft 365 Copilot Chat, in ways that carried patient-safety risk, and that VHA had no formal mechanism to identify, track, or resolve the risks those tools introduced. The final report followed in June 2026 with three recommendations. VA's Under Secretary for Health concurred, with a target completion date of April 2027.
Hold the scale of what was running while that gap was open. VA's own 2025 AI use case inventory documents 367 AI use cases across the department. VA GPT alone reports more than 95,000 users. The tools were deployed at scale first, and the mechanism to govern them, to log their use, monitor their accuracy, and resolve their failures, is the thing VA has until April 2027 to build.
The sequence is the story. Deployment moved faster than assurance, an inspector general documented the gap, and the department agreed and set a date to close it. That date is a written, mandated requirement for audit logging, model monitoring, and post-deployment review. It is a governance market with a deadline attached, and it exists because the model was the easy part and the accountability around it was not.
The sleeper is the encryption that has to outlive the patient
The layer furthest down the stack is the one with the longest fuse, and it lands hardest on health data.
On June 22, 2026, the President signed two quantum executive orders. One, EO 14413, funds the next generation of quantum hardware. The other, EO 14412, accelerates the migration to post-quantum cryptography, the encryption designed to survive a quantum computer capable of breaking the encryption in use today. That order sets hard dates: key establishment by the end of 2030, digital signatures by the end of 2031, and a Federal Acquisition Regulation rule, due within 180 days, requiring covered contractors to meet NIST's post-quantum standards. NIST finalized those standards, FIPS 203, 204, and 205, back in August 2024. The government's own estimate to migrate its prioritized civilian systems runs about $7.1 billion through 2035.
Federal health agencies are first in line for a specific reason. They hold decades of protected health information, records that stay sensitive for a veteran's entire life. That makes them the prime target for what cryptographers call harvest now, decrypt later: an adversary copies the encrypted data today and waits for the machine that can open it. A stolen credit card expires. A stolen medical history does not.
Here is the gap in the record, stated as a fact rather than an alarm. As of the middle of 2026 there is no guidance from HHS or its Office for Civil Rights linking post-quantum cryptography to HIPAA. The one rule that would have forced the issue, the 2025 proposed update to the HIPAA Security Rule that would make encryption of electronic health information mandatory, missed its targeted May 2026 finalization and remains a proposal. So health systems face the prospect of rebuilding their encryption once to satisfy a coming HIPAA mandate, and again to satisfy the post-quantum deadline, because no federal body has yet told them to do both at the same time. Migrate once and you have solved the problem. Migrate twice and you have paid for it twice.
The model you name has to be able to run where the contract does
One more layer sits under every AI-bearing proposal, and it lives in the fine print rather than the tech stack: the vendor's own terms of use.
When the Department of War cleared eight firms for classified networks on May 1, 2026, one name that many expected was absent. Anthropic was excluded, after declining to accept "any lawful use" terms. The detail worth noticing is the inversion underneath it. Anthropic's civilian authorization, FedRAMP High plus Impact Level 5, actually exceeds OpenAI's FedRAMP Moderate. The company with the stronger civilian accreditation was the one kept off the classified network, because accreditation level is necessary but not sufficient, and a vendor's own acceptable-use policy can be the disqualifier.
The operational lesson is narrow and worth carrying. A model's benchmark score does not decide whether it can legally run where your contract requires. Its authorization level and its published use terms do. For a civilian VA, CMS, or HHS pursuit, a High-plus-IL5 model is fully usable and arguably the stronger choice. For a classified defense pursuit, the same model may be off the table. Confirm the model can run where the work lives before you write it into a proposal, not after.
China wired the rules to the deployment; we did the reverse
The country this entire race invokes is running the same technology on a different sequence, and the contrast is the clearest way to see what the United States has chosen.
China moved payment and ethics rules into place alongside deployment rather than behind it. Its National Health Commission published a reference guide of AI application scenarios in late 2024, followed by implementation opinions in November 2025 and, in December 2025, a decision by the national health-insurance administration to classify AI-assisted diagnosis as a billable, reimbursable item. In April 2026 it issued ethics-review measures requiring independent committees for medical AI. Its first fully integrated AI hospital opened in Boao, in Hainan, in March 2026. The deployment figures are contested, ranging from roughly ninety tertiary hospitals in the peer-reviewed literature to at least three hundred in journal reporting, and a good share of the underlying studies are low quality. The direction is not.
The structural difference is the whole point. China wired reimbursement and ethics to deployment. The United States deployed at scale first, through a dollar-a-seat schedule and a governance layer it is now building behind the fact. Neither approach is obviously right. But they produce different bills, and the American bill is the one coming due in the layers under the model.
What the dollar actually bought
Stack the pieces and the shape is plain. The frontier model entered federal health at a dollar and stopped conferring an advantage the moment the government bought four of them. The money did not disappear. It moved to the platform the model rides on, to the integration work DHA just pulled back onto the open market, to the governance mechanism the VA OIG gave the department until 2027 to build, and to the encryption migration no health agency has yet been told how to sequence. Those are the layers a health-IT integrator already owns. They are also the layers almost no one priced into the AI story, because the story was about the model, and the model turned out to be the cheap part.
None of this is a race that federal health is losing. It is a race whose scoreboard is in the wrong place. The frontier layer is settled and commoditized. The contest that decides whether any of this reaches a patient safely runs one layer down, in accreditation, governance, integration, and cryptography, and that contest is barely started.
Which brings it back to the person at the end of the wire. A veteran walks into a VA clinic. Her clinician dictates the visit into an ambient scribe that drafts the note, a tool the inspector general has said runs without a mechanism to catch its own mistakes. That note joins a record that will stay sensitive for the rest of her life, protected by encryption a future computer is expected to break, under a HIPAA rule that has not yet told anyone to fix it. The dollar bought the model that wrote the note. Everything that keeps her record safe, accurate, and accountable is the part still under construction. That is where the mission is now. That is where the market is too.
Let's roll.
— Mary
Mission Meets Tech
The views expressed in this newsletter are my own and do not represent the official position of any organization. This content is for informational purposes only.
MMT Premium
This issue maps where the value moved. The companion Capture Corner maps how to reach it. Premium subscribers get the four capture lanes the roll-down just opened: the MHS GENESIS integrator unwind with the sole-source notice ID, the incumbent vehicles, and the transition clock; the VA AI governance-and-audit market the OIG mandated, with the buyer to watch and the three-recommendation brief that sells it as risk closure; the post-quantum-to-HIPAA advisory lane with no incumbent and the GSA routes that actually buy it; and the vendor-use-policy gate that turned a teaming assumption into a disqualifier, with the FedRAMP and Impact Level matrix a capture lead needs before naming a model in a proposal.
Founding Member rate: $199/year, locked permanently for the first 100 subscribers.
Standard rate: $249/year or $29/month.
Premium adds 48-hour early access to deep-dive analysis, monthly Capture Intelligence Sheets with sourced action windows, direct Q&A access (reply to any premium issue), and tool discounts: ProposalPulse $14.99 per assessment, MarketPulse $35 per brief.
Subscribe at missionmeetstech.com/pricing.
Sources
OneGov pricing and scale: GSA, "GSA Announces New Partnership With OpenAI" ($1/agency, Aug. 6, 2025); Anthropic, "Offering Expanded Claude Access Across All Three Branches of Government" ($1, Aug. 12, 2025); GSA Buy AI (Gemini for Government $0.47, Aug. 21, 2025); GSA, "GSA, xAI Partner" (Grok $0.42, Sep. 25, 2025). · Nineteen agreements / up to 90% discount: GSA End-of-Year OneGov Update (Dec. 30, 2025). · Twenty companies, 120+ orders, ~3.4M users, ~$1.1B first-year savings: Nextgov (June 2026). · Claude for Government standard rate $60/seat/month: Anthropic Public Sector FAQs. · Department of War cleared eight firms for classified networks (AWS, Google, Microsoft, OpenAI, SpaceX, NVIDIA, Reflection, Oracle), Anthropic excluded on use-policy grounds, May 1, 2026: Breaking Defense; DefenseScoop. · OpenAI FedRAMP Moderate vs. Anthropic Claude for Government FedRAMP High + IL5: OpenAI, "OpenAI available at FedRAMP Moderate"; Claude for Government FAQ. · VA ambient scribe: Rise8 (SDVOSB) with Thoughtworks, contract 36C10B26C0025, $4,945,717.92, sole-source, awarded Feb. 19, 2026, engines Abridge and Knowtex (HigherGov; orangeslices.ai). · VA EHRM: VA internal estimate $37.2B (Sept. 2025), ~$12.7B spent, Accenture Federal Services integrator, deployments resuming FY2026 (MeriTalk; Healthcare Dive). · DHA Clinical AI Agent: six-week limited fielding Oct. 31–Dec. 11, 2025 (Madigan, Camp Lejeune, Portsmouth, Wilford Hall), phased rollout from Feb. 2026; ~1.2 hrs/day charting reduction, 26% after-hours reduction at Camp Lejeune; competed via DHMSM CSO Area of Interest #3, solicitation HT0038-25-X-0003, as OT agreement under 10 U.S.C. 4022 (dha.mil; Navy Medicine; HigherGov). · MHS GENESIS baseline: DHMSM award to Leidos $4.3B (2015), $5.5B after 2018 modification (GAO-22-104521; DODIG-2016-094). · DHA integrator unwind: Leidos LSI role ending; five sole-source FFP outcome-based platform awards (Oracle Health, Philips, Amwell, Henry Schein, Solventum) with transition dates July 2026–July 2027; services layers competed full-and-open to eliminate legacy integrator pass-through costs; SAM.gov notice opp 8eab9996b3eb4632ace81014f963c351 (Washington Technology, June 2026; SAM.gov). · VA OIG: preliminary advisory memorandum 26-00182-42 (Jan. 15, 2026) finding patient-safety risk from clinical use of VA GPT and Microsoft 365 Copilot Chat and no formal mechanism to identify, track, or resolve gen-AI risk; final report 26-00182-140 (June 2026), three recommendations, USH concurred, April 2027 target (vaoig.gov; FedScoop). · VA AI inventory: 367 AI use cases, VA GPT 95,000+ users (department.va.gov/ai). · Quantum EOs: EO 14413 and EO 14412, signed June 22, 2026; EO 14412 key establishment by Dec. 31, 2030, signatures by Dec. 31, 2031, FAR rule within 180 days for covered contractors; OMB M-26-15 (June 24, 2026), agency plans within ~120 days (whitehouse.gov). · NIST FIPS 203/204/205 finalized Aug. 13, 2024 (csrc.nist.gov). · ~$7.1B civilian PQC migration estimate, 2025–2035 (White House July 2024 report). · No HHS/OCR guidance links PQC to HIPAA as of mid-2026 (documented absence). · HIPAA Security Rule NPRM (90 FR 898, Jan. 6, 2025) mandating ePHI encryption missed targeted May 2026 finalization, remains proposed. · China healthcare-AI policy sequence: NHC AI application scenarios reference guide (Nov. 2024); "AI + Medical and Health" implementation opinions (Nov. 4, 2025); NHSA classification of AI-assisted diagnosis as billable (Dec. 2025); AI ethics-review measures (Apr. 3, 2026); Boao AI hospital (Hainan, Mar. 26, 2026) (Chambers Digital Healthcare 2026 – China; Global Times/Xinhua). · China DeepSeek hospital adoption: ~90 tertiary hospitals with two-thirds of studies low quality (JMIR e93354); at least 300 hospitals in clinical diagnostics by early March 2025 (JAMA via SCMP). · FY2027 AI-related defense spending scaling toward ~$58.5B (Greenberg Traurig FY2027 analysis).
