VA’s Enterprise Cloud was a real accomplishment. Hundreds of applications migrated. VistA, all 130+ instances of it, running in the cloud. The foundation is solid.

The question is: what should VA build on top of it?

An AI Platform Built to Serve Every Veteran

The next step should be a shared intelligence layer — something every VA program, every system, and every one of the 17.6 million veterans that VA exists to serve can draw from.

Not a chatbot bolted onto a portal. Not one of 367 AI pilot programs running in sandboxes. A purpose-built, enterprise-grade AI backbone that would give VA something it has never had: a single place where all of its data makes sense together, where any policy question would have an authoritative answer, and where every veteran would have a tireless digital advocate working on their behalf.

What the AI Backbone Would Actually Do

At its core, the AI backbone would do three things — and they are interdependent.

First, it would give VA’s AI a coherent view of its own data estate. Today, VA’s data lives in hundreds of systems: VistA, VBMS, MBMS, CorpDB, GI Bill, Home Loan Guaranty, and more. Each is authoritative for something. A few talk fluently with others. No AI can reason across that estate today because no single layer connects it. The backbone would change that — continuously reconciling data across sources, surfacing conflicts rather than burying them, and making the full picture queryable. When VA data says one thing and Pentagon personnel records say another, the backbone would flag that discrepancy rather than let it silently corrupt a veteran’s benefit picture downstream.

Second, it would make VA policy accessible, legible, and usable at the point of decision. The backbone would make VA’s full policy corpus — CFR Title 38, the M21-1 Adjudication Manual, Board of Veterans Appeals decisions, IG and GAO reports — navigable in real time, with clear sourcing, so that answers are not only fast but defensible. The institutional knowledge that currently lives in the heads of experienced VSOs and senior claims processors would be available to everyone, including the veteran who doesn’t have a VSO.

Third, and most importantly, it could give every veteran a persistent digital advocate. Every living veteran could have a persistent AI agent — personalized to their service history and benefits profile — that would proactively track their entitlements across every VA program and assist the veteran without being asked. When a new law passes that changes eligibility, the twin would know. When a veteran’s circumstances change, the twin would adapt. When a benefit is going unused, the twin would surface it.

What the Compute Architecture Might Look Like

This is not a software project. It is a physical infrastructure build, and the scale is serious.

I would place the primary AI compute cluster at VA’s Austin Information Technology Center — AITC Austin. This is already VA’s major enterprise data center with the physical security posture and network backbone required for this AI workload. At initial operating capability, you’d perhaps be looking at 1,000 high-density GPU nodes configured in a purpose-built AI compute cluster with InfiniBand network fabric between nodes. Storage could run into the petabytes of mixed NVMe and object storage. Model weights alone could exceed a terabyte. The policy corpus, vector stores, and 17.6 million veteran twin memory stores add up fast.

I would have VA’s Martinsburg, West Virginia data center serve as a warm standby — a smaller mirror cluster providing geographic redundancy and disaster recovery.

At the facility level — major VAMCs, Regional Offices, and call centers — smaller inference appliances could handle latency-sensitive workloads. Edge nodes sized for inference only could handle real-time sessions locally and sync state back to the primary cluster. Think mini-AI backbones deployed like VistA instances once were: locally present, centrally governed.

A sovereign cloud overflow tier would live inside VA’s existing VAEC for burst capacity when legislative or catastrophic weather events trigger simultaneous processing across millions of twin profiles. This architecture could draw 10 to 20 megawatts at full load — a serious conversation about power and cooling infrastructure at Austin and Martinsburg. Tough but well-understood engineering problems with known solutions.

Avoiding the Next Failure Mode

VA’s VASI catalogs roughly 1,200 IT systems across VHA, VBA, NCA, and VA’s staff offices. Every major software vendor has embedded AI capability — and if used properly, that’s a good thing. But the introduction of AI into each of those systems is also likely to create a new kind of fragmentation: different models, different data views, different answers to the same question, and little visibility into how decisions are being supported across VA’s mission.

AI entropy will prevail — inconsistent outputs, conflicting data sources, redundant spend — unless VA gets ahead of it. The backbone would change that dynamic. It would function as the AI layer behind these systems. Each system, whether a custom MUMPS application from 1989 or a modern VBMS instance, could call the backbone’s APIs to access model inference, retrieval from the enterprise data fabric, policy engine queries, and veteran twin context — without building any of that capability itself.

For the hundreds of legacy systems, this would mean AI-augmented capabilities through a secure API call rather than a multi-year modernization program. Build it once. Make it available to everything.

Built for Everyone Who Served

As of March 2026, the projected U.S. veteran population is 17.6 million. Of those, 8.62 million are enrolled in VA health care — meaning more than half of America’s living veterans have no active relationship with the VA healthcare system. Yet VA’s benefits portfolio extends far beyond healthcare.

6.5 million veterans receive VA disability compensation. 2 million are rated 100% disabled. Nearly 950,000 used VA education benefits in FY2025. 4 million are active VA home loan participants. 544,636 surviving spouses receive Dependency and Indemnity Compensation. VA supervises 5.55 million life insurance policies with a face value of $1.5 trillion.

A veteran rated 100% disabled who is also a surviving spouse of a prior marriage, using Chapter 31 vocational rehabilitation, with an active VA home loan, and approaching Medicare eligibility — that veteran’s benefit picture is not something any single VA employee, VSO, or portal was designed to handle easily. A digital twin would hold it all at once. The marginal compute cost of the ten-millionth twin would be close to zero once the platform is running. Build it at full scale from day one, and you would have a platform worthy of the obligation.

How You Build Something Like This

The platform would live inside VA’s IT security boundary — the same boundary that already governs VA’s Azure and AWS cloud environments. Data stays inside that boundary. Processing happens inside that boundary. VA’s governance, VA’s security controls, VA’s data rights.

I’m not suggesting VA build this. The technology to do all three exists commercially today. VA should subscribe to this AI backbone through a fixed-price, Contractor-Owned, Contractor-Operated platform operating inside the VA, using performance-based SLAs — an OPEX approach, not CAPEX. VA has spent decades accumulating IT assets it cannot afford to maintain. The AI backbone should be different from day one. The contractor would bear the risk and capital cost. VA would pay for outcomes, not effort.

One thing VA should not do: follow the Pentagon’s GenAI deployment model. VA will need model-agnostic layers. Vendor-agnostic platforms. The discipline to abstract the model from the application so that when a better model emerges — and in this market, better models will emerge constantly — VA can swap it in without rebuilding every system that depends on it. Build to the interface, not the implementation.

VA Built the Foundation. Time to Build the Intelligence Layer.

VA’s cloud migration was a genuine engineering feat — one I’m proud to have helped launch. Now it’s time to step forward again. Build what’s never existed: a single, shared, continuously improving intelligence layer that would know everything VA knows, that would speak to every veteran regardless of which program they’ve used, and that would work on their behalf without requiring them to become experts in the system.

VA built the cloud. Now it needs the brain.