Most AI deals announced these days are about models, benchmarks, or billion-dollar funding rounds. The one worth watching this season is none of those. It is a procurement agreement — the kind of paperwork that usually puts people to sleep — between the US General Services Administration (GSA) and the data-cloud company Snowflake. And it may tell us more about how artificial intelligence gets priced over the next decade than any model launch.
The reason is simple: when a buyer the size of the entire US federal government negotiates a discount, it doesn’t just save money. It sets a reference price for everyone else.
The deal
Under a so-called OneGov agreement, the GSA has made Snowflake’s AI and cloud-based data products available to all federal agencies at negotiated discounts. According to terms reported by Crescendo AI (2026, pending verification against GSA’s own publication), agencies get roughly 20% off compute — a discount that scales up to 50% as usage rises — and about 27% off storage.
The structure matters as much as the headline numbers. A flat discount is a one-time saving; a usage-scaling discount is an incentive. The more agencies consolidate their data and workloads onto the platform, the cheaper each unit becomes. That design nudges federal departments toward exactly the behaviour the deal is meant to encourage: pulling fragmented data out of dozens of incompatible systems and into a common analytical layer.
That is the stated intent. The roughly one-year-old OneGov framework is designed, per Crescendo AI, to help federal agencies eliminate data silos, modernize ageing IT infrastructure, and improve mission effectiveness. In plain terms, government data tends to sit in disconnected pools — one for benefits, one for tax, one for procurement — that rarely talk to each other. Pooling buying power to standardize the tooling is meant to break those walls down faster and at lower cost than each agency negotiating alone.
Why bulk procurement matters
Here is the part that should interest anyone outside Washington. AI pricing is still being figured out in real time. Compute costs, token costs, storage costs — none of them have settled. In that fluid environment, a buyer large enough to move volume doesn’t just take the market price. It helps set it.
When the US federal government negotiates 20-to-50% off compute, it establishes a benchmark that state governments, large enterprises, and even other national governments will point to in their own negotiations. “The feds got this rate” becomes a bargaining chip. Public-sector buying power, in other words, shapes AI pricing for the whole ecosystem — a quiet but genuine lever of industrial policy.
There are real upsides for the agencies themselves:
- Speed. A pre-negotiated framework means an agency can adopt modern data and AI tooling without running its own months-long procurement from scratch.
- Standardization. Common platforms make it easier to share data, hire skills, and build once-and-reuse systems across departments.
- Cost discipline. Volume discounts and usage tiers reward consolidation and discourage redundant spending.
But standardization is a double-edged sword, and the second edge is lock-in. The same consolidation that delivers savings and interoperability also concentrates a government’s most critical data and workloads on a single commercial platform. The trade-off between standardization and dependency is the central tension in every deal of this kind — and it is rarely resolved in the press release.
The risks
The first risk is concentration. OneGov-style agreements, by design, funnel demand toward a small number of pre-approved vendors. That is efficient. It is also how a competitive market quietly narrows into an oligopoly of preferred suppliers. If the bulk of federal data and AI workloads gravitate to a handful of platforms, future negotiating leverage erodes — the buyer becomes dependent on the very vendors it once squeezed.
The second risk is data governance and portability. Discounts that scale with usage are, functionally, discounts that reward you for moving more of your data onto the platform and keeping it there. The cheaper it gets to stay, the more expensive — in time, engineering, and egress fees — it becomes to leave. Governments hold the most sensitive datasets in any society: tax records, health data, citizen identity. The terms governing who can access that data, where it physically resides, and how cleanly it can be exported should carry as much weight as the headline discount. Often they don’t.
The third risk is dependency dressed up as modernization. “Eliminating data silos” is a worthy goal. But replacing many silos with one large, proprietary silo is not the same as building genuinely open, portable infrastructure. The healthiest version of these deals insists on open formats, documented exit paths, and the ability to run a second vendor without re-architecting everything. Buyers with real leverage — and a government is the buyer with the most leverage of all — should spend some of that leverage on portability, not just price.
The India read
For Indian policymakers, the GSA-Snowflake model is less a deal to copy than a template to study. India already has the institutional plumbing for this kind of aggregated buying: the Government e-Marketplace (GeM) has spent years standardizing public procurement, and central cloud arrangements like MeghRaj have tried to consolidate government IT. What India has not yet done at scale is point that combined buying power squarely at AI and data infrastructure.
The opportunity is significant. India’s public sector — Union ministries, state governments, public-sector undertakings, and the digital-public-infrastructure stack that powers everything from payments to identity — represents enormous, predictable demand. Aggregating even a slice of that demand into a single negotiated framework could make AI and data tooling dramatically more affordable at public scale, and set price benchmarks that ripple out to Indian startups and enterprises too.
But the risks travel with the model. India should take the GSA playbook’s strengths — pre-negotiated rates, usage incentives, faster modernization — while explicitly correcting for its weaknesses:
- Mandate diversity. Structure frameworks so that two or more vendors can serve the same need, preserving competition and negotiating leverage rather than crowning a single winner.
- Buy portability, not just price. Make open data formats, documented export paths, and capped egress costs non-negotiable contract terms — using the discount conversation as the moment to extract them.
- Protect sovereignty. For citizen-scale datasets, data residency, access controls, and auditability matter more than the size of the discount. Cheap compute is no bargain if it compromises governance.
- Favour the stack India already has. Where domestic and open-source options can meet the need, aggregated buying power should be used to strengthen them, not to deepen dependence on a single foreign platform.
The deeper lesson cuts across both countries. The era in which governments quietly accepted whatever AI vendors charged is ending. The GSA’s agreement is an early signal that public buyers can — and increasingly will — use their sheer scale to bend AI prices downward. The question for India, and for every public buyer watching, is whether they use that leverage only to get cheaper bills, or whether they use it to demand something harder to win once the contracts are signed: openness, portability, and the freedom to walk away. Price is the easy victory. Independence is the one worth negotiating for.
