The IRS is buying an Nvidia AI supercomputer

The IRS is preparing to bring serious AI horsepower to tax collection.

According to an unannounced procurement document dated February 5 and reviewed by The Intercept, the agency plans to purchase a state-of-the-art Nvidia SuperPod computing cluster. The setup will be housed at the IRS computing center in Martinsburg, West Virginia, combining 31 separate servers packed with 248 flagship Blackwell chips and a substantial memory upgrade.

While the hardware has not been officially purchased yet and carries no listed price tag, SuperPod setups typically start at around $7 million. For scale, federally funded research lab MITRE spent $20 million last year on a similar setup to build custom models for federal agencies.

Turning tax returns into training data

The IRS has not officially disclosed what the machine will do or even which presidential administration authorized the order. A Treasury Inspector General report from 2024 listed 68 active AI projects across the agency, though many details were redacted.

The procurement text itself offers clear hints. It notes that the hardware will be managed by the agency's Research, Applied Analytics, and Statistics division (RAAS) to handle complex machine-learning workloads. That division's Compliance Data Warehouse team previously used AI for fraud detection, identity theft prevention, and tracking taxpayer behavior.

As tax attorney Travis Thompson pointed out, the IRS sits on an unprecedented mountain of untapped financial records. Training specialized internal models on that proprietary archive requires serious compute, and an on-premise Blackwell cluster fits that need perfectly.

DOGE, automation, and shrinking the bureaucracy

The timing lines up with a massive push from the Trump administration and Elon Musk's Department of Government Efficiency (DOGE). DOGE personnel have recently arrived at the IRS to evaluate IT infrastructure, explicitly seeking to replace human staff with automated software.

This trend has been building for a while. A June 2024 Government Accountability Office report urged the agency to use AI tools to help track down hundreds of billions of dollars in unpaid taxes each year.

Thompson noted that cutting federal headcounts and aggressively deploying machine learning go hand-in-glove. If the administration wants to shrink agency personnel while closing the tax gap, automated systems have to shoulder the burden.

Why it matters

For everyday taxpayers and businesses, an IRS powered by top-tier AI means audit selection will get far more granular. Sophisticated tax shelters that previously flew under the radar due to staffing shortages could now be flagged in seconds by bespoke models trained on decades of filings.

From an infrastructure standpoint, buying a dedicated, on-premise SuperPod is the right move for handling sensitive financial data. The agency cannot simply pump confidential returns into public commercial clouds. Having an in-house compute island solves the immediate data sovereignty and privacy concerns.

However, there is a serious catch. Rushing to replace human auditors with algorithmic workflows risks escalating false positives and leaving ordinary taxpayers trapped in automated disputes with no human rep to talk to. Using Blackwell chips to hunt down major fraud makes total sense; using them as an excuse to gut human oversight could backfire fast.