Anthropic on Tuesday announced that it will make its Claude for Enterprise and Claude for Government platforms available to all three branches of the U.S. government — executive, legislative, and judicial — for just $1 per agency per year.

The offer, made in partnership with the U.S. General Services Administration, will also include technical support to help agencies integrate the company’s AI tools into operations.

The move mirrors a similar step by rival OpenAI, which earlier this month revealed it would provide its ChatGPT Enterprise product to U.S. federal agencies for $1 over the next year. Both companies are offering these steeply discounted terms as part of an aggressive push to deepen ties with policymakers and regulators, a strategy that analysts say could influence future AI governance frameworks.

“America’s AI leadership requires that our government institutions have access to the most capable, secure AI tools available,” Anthropic CEO Dario Amodei said in a statement.

The competition for U.S. government customers has become increasingly fierce as major AI players see federal adoption not only as a lucrative long-term market but also as a powerful endorsement of their technology’s credibility and security. In June, Anthropic rolled out a set of Claude Gov models built exclusively for U.S. national security applications. Just one month later, the Department of Defense awarded contracts worth up to $200 million to Anthropic, Google, OpenAI, and Elon Musk’s xAI for AI development. On the same day, Musk’s xAI announced Grok for Government, a suite of tools aimed at federal agencies.

OpenAI, for its part, has been expanding its physical presence and policy footprint. The company plans to open its first Washington, D.C., office early next year and launched its OpenAI for Government program in June, underscoring its focus on the public sector.

Many believe that Anthropic’s latest move may be aimed at cementing its position as a trusted AI partner for the U.S. government before rival firms can lock in longer-term contracts. Federal technology procurement often favors established vendors, and an early presence in multiple branches could give Anthropic a lasting advantage.

The $1 pricing model is symbolic rather than profit-driven, signaling that the real prize lies in influence, long-term service agreements, and the potential integration of AI into critical government workflows.

The battle for government adoption is also occurring against a backdrop of heightened AI regulation debates in Washington. Companies that can position themselves as cooperative partners in safe and secure AI deployment may have more sway in shaping emerging policy frameworks.

At the same time, the Pentagon’s funding commitments to multiple AI developers show that the U.S. government is actively fostering competition in the sector rather than relying on a single supplier. For firms like Anthropic, this means differentiation through product performance, security assurances, and adaptability to classified and sensitive environments will be key.

Washington has increasingly turned to AI companies for critical projects, from enhancing national defense capabilities to automating bureaucratic processes.

Defense-related AI has been a key driver in this trend. The Pentagon has been exploring AI for battlefield intelligence, logistics optimization, and cyber defense, while civilian agencies are using AI to improve data processing, fraud detection, and service delivery.

For companies like Anthropic, OpenAI, Google, and xAI, securing these contracts offers long-term strategic benefits, including privileged access to government datasets, opportunities to test AI in high-stakes environments, and an elevated role in shaping regulatory frameworks.

With President Donald Trump’s administration signaling a strong commitment to advancing U.S. AI dominance, analysts expect the fight for federal AI contracts to intensify in the months ahead.

Huawei has unveiled Unified Cache Manager (UCM), a software layer designed to speed up large-model inference by moving data across HBM, DRAM, and SSD according to each workload’s latency needs.

Company executives in Shanghai said lab testing showed latency cuts of up to 90% and throughput gains as high as 22x. Huawei plans to open-source UCM in September, first to its developer community and then industry-wide.

The pitch is straightforward: if software squeezes more performance out of commodity memory, Chinese providers can deliver competitive AI inference without leaning as heavily on scarce, expensive high-bandwidth memory (HBM). That matters because the global HBM market is surging—about $34bn this year on a path toward $98bn by 2030—and supply is dominated by SK Hynix, Samsung, and Micron, all outside China’s control.

UCM arrives as Beijing accelerates its push for chip self-reliance. Local memory and packaging players—Yangtze Memory Technologies, Changxin Memory Technologies, Tongfu Microelectronics—are still scaling toward HBM2, while leading foreign rivals are already commercializing HBM4. Limited access to advanced tools and production equipment under allied export regimes continues to slow the catch-up.

Why UCM matters now

The timing intersects with two pieces of U.S.–China semiconductor drama. First, Nvidia’s H20, a downgraded AI accelerator tailored for China after Washington tightened restrictions, has been cleared for licensed sales under a 15% revenue-sharing arrangement with the U.S. government. Second, President Donald Trump has said he’s open to a deal that would allow Nvidia to ship a scaled-down Blackwell to China, provided performance is cut by 30% to 50%.

UCM could weigh on both paths. If inference efficiency improves meaningfully on mixed-memory systems, Chinese buyers have a stronger case to delay or downsize H20 purchases, especially while security questions hang over that product in China’s public discourse. And if UCM gains traction across domestic stacks—particularly on Huawei Ascend systems paired with the CANN software toolkit and new CloudMatrix 384 “supernode” boxes—it reduces the urgency to lobby for a Blackwell variant. In short, the more value China can wring from software and local silicon, the less leverage Washington gains from gated access to U.S. top-end chips.

How we got here: a concise timeline of restrictions and responses

Over the past three years, export policy and industry workarounds have moved in lockstep. Here’s the arc that ties Huawei’s UCM launch to today’s licensing and tariff chessboard—told as a sequence rather than a list, to preserve the flow of events.

In late 2023, Washington tightened rules on advanced AI accelerators bound for China and also targeted the most capable HBM configurations. U.S., Dutch, and Japanese toolmakers aligned on curbs for leading-edge lithography and related equipment, making it harder for China to jump nodes or scale high-bandwidth memory at pace. Chinese firms doubled down on systems-level innovation, compiler work, sparsity, quantization, and model engineering to stretch the chips they could still buy.

By 2024, Chinese labs and startups showed credible results on constrained hardware; DeepSeek became a reference point for aggressive software optimization on limited compute. Huawei, already under U.S. sanctions, pushed its Ascend roadmap, accelerated work on CANN, and began previewing larger AI clusters—an overt play to build a rival software ecosystem to Nvidia’s CUDA while assembling domestically sourced platforms.

In early 2025, the U.S. signaled even tighter guardrails on top-tier accelerators and high-bandwidth memory, while back-channel trade talks opened space for case-by-case licensing. Beijing pressed for relief on HBM specifically, arguing that inference build-outs need memory bandwidth more than peak FLOPs. Washington floated selective permissions tied to usage, end customers, and performance ceilings, setting the stage for bespoke arrangements.

Through mid-2025, Nvidia’s H20—a slowed, China-specific Hopper derivative—became a policy test case: first barred, then licensed, and finally allowed under a 15% revenue-for-license deal with the U.S. government. At the same time, state-affiliated Chinese outlets questioned H20’s safety and value, urging buyers to consider domestic options for sensitive workloads. The result was a mixed demand signal: some Chinese firms weighed H20 for near-term capacity, others pivoted to domestic stacks and software routes that promised acceptable performance per dollar without foreign-policy risk.

As summer turned to Q3 2025, President Trump publicly said he would consider a scaled Blackwell for China—again contingent on performance downgrades—and confirmed the 15% take on licensed H20 shipments, with AMD’s MI308 on the same footing. That real-time bargaining underscored a trade posture that ties market access to direct fiscal returns and performance controls, even as national-security framings remain in place.

The 15% revenue share demanded on H20 and MI308 licenses is unprecedented in the history of U.S. export controls, which traditionally rely on entity listings, end-use checks, and outright performance thresholds rather than ongoing fiscal participation. It signals a turn toward transactional licensing—permission as a meter that can be dialed up or down by product class, customer set, or geopolitical moment. For companies, it creates a variable tax on a key market, alters margin calculus, and invites questions about whether similar levies could spread to other domains, from data center components to advanced manufacturing tools.

That approach may also complicate the national-security rationale. If a product is risky enough to restrict, critics ask, why is it safe once a revenue cut is paid? The counterview is that the levy functions as both a deterrent and a data point, giving Washington visibility and leverage while preserving some influence over China’s AI stack.

What Huawei’s UCM changes in practice

If UCM’s gains hold up outside Huawei’s labs, integrators can design inference clusters that rely less on the latest HBM, substituting cheaper DRAM and SSD while meeting service-level targets through smarter caching and prefetch. That favors domestic suppliers in three ways. First, it lessens the immediate need for foreign HBM supply. Second, it improves the economics of Ascend-based deployments where memory is a bottleneck. Third, it strengthens the CANN software ecosystem by anchoring it to a tangible performance win that developers can adopt quickly once UCM is open-sourced.

This introduces two near-term risks in China for Nvidia. Licensed H20 demand could soften if buyers judge that software-optimized domestic stacks are “good enough” for mainstream inference, especially where security reviewers prefer local platforms. And any future scaled-down Blackwell deal might face a tougher value proposition if UCM enables competitive latency and throughput on non-HBM-heavy designs.

The summit track and the chip ledger

Trade negotiators are working toward a possible Trump–Xi meeting. Beijing has made HBM relief a priority ask, arguing it unlocks capacity without handing over leading-edge compute. Washington has shown interest in license-and-limit frameworks, plus the new revenue take.

UCM’s entrance strengthens China’s negotiating position: it adds a credible Plan B that reduces reliance on foreign memory and erodes the leverage that comes with scarce HBM supply. That, in turn, could push the U.S. side to seek broader concessions—on transparency, on end-use audits, or on carve-outs tied to specific operators—in exchange for any HBM easing.

Three signposts will tell us whether UCM reshapes this market or merely trims costs at the margin. First, adoption velocity once the code is released, and whether top Chinese cloud providers standardize it across Ascend and x86/GPU fleets. Second, procurement patterns for H20 through year-end, including any cancellations or substitutions in government-adjacent sectors. Third, the language that emerges from any Trump–Xi agreement on HBM and export licenses, especially clauses that link performance ceilings to ongoing fiscal terms.

However, Huawei has put another software lever on the table for now. If it performs as advertised, the balance of power in China’s inference build-out tilts a little further from hardware choke points and a little closer to systems engineering—exactly where Beijing wants it.