Lead Analysis
Strategy6 min

Claude Enters GA on Microsoft Foundry Regarding Blackwell Ultra, Closing the Azure-AWS-Google Triangle

Engenheira de plantão em corredor de data center diante de racks NVIDIA GB300 com LEDs azuis acesos.

Anthropic releases Opus 4.8 and Haiku 4.5 in production on Azure running on NVIDIA GB300 chips. Availability kicks off in East US and West Europe; UK South and Asia will follow by mid-July.

Anthropic made Claude generally available on Microsoft Foundry on Monday (29), concluding the preview period that started in May. The infrastructure is the NVIDIA GB300, Blackwell Ultra generation, which the manufacturer began to ship in volume starting in the second quarter. Claude Opus 4.8 and Claude Haiku 4.5 are already listed on the Messages API hosted by Microsoft, with inference running in East US and West Europe and planned expansions to Southeast Asia, Australia East, and UK South by mid-July.


Microsoft cites a 40% increase in token generation speed compared to H100 clusters and a 15% increase over B200 in the benchmarks released alongside the GA. Entry prices start at $0.25 per million tokens for Haiku, $3 for Sonnet, and $15 for Opus, mirroring the pricing table of Anthropic’s own API. What changes in the Foundry is the commercial framework: Claude is now part of the same Enterprise Agreement the company already uses for Microsoft 365, Defender, and the OpenAI catalog, with unified billing, IAM, and auditing.


What Changes for the CIO and CISO


For the CIO who is already purchasing Azure, the GA removes two friction points. The first is regulatory: the option for a US data zone allows sensitive data to remain confined to American regions, a crucial point that had been blocking pilots in European banks operating under heavy GDPR and DORA compliance. The second is the flattening of governance, because Defender for Cloud apps now recognizes traffic to Claude as just another endpoint of the Foundry, without needing to redo DLP policies created for GPT-5.5.


For the CISO, the assessment is more complex. Inference on GB300 opens a useful window of context that fits without severe degradation, bringing the feasibility of agents orchestrating long flows within the SOC closer to reality. However, the governance surface expands: each Foundry client will have to deal with the possibility of Claude agents executing code against private MCP servers, and the controls that Anthropic announced parallelly, featuring self-hosted sandboxes on providers like Cloudflare, Modal, and Vercel, acknowledge this concern. Those operating strict Zero Trust will need to tie this back to their PAM rather than delegating it to the agent loop.


The Geopolitics of Compute


The British and Australian trail for July deserves attention. Nine months ago, the standard reading in the industry was that regulated workloads in the UK would have to wait for the national infrastructure of hyperscalers to mature. The updated timeline now pulls the local IT services providers to reorganize earlier. PwC’s Acceleration Center in Bangalore, which has been piloting Claude since April within the Microsoft framework, gains a regional counterpart in the UK without needing to rewrite the data pipeline, and HCLTech has already notified clients in the insurance sector that onboarding teams to the Foundry will commence in July.


In Brazil, the expansion reaches two systems that were procuring the product outside of the preferred terms. Itaú announced in May that it would standardize Claude in a back office front, and Bradesco hired Capgemini for an equivalent front. Both will migrate to the Foundry without renegotiating processing terms, reducing pressure on the IT cost centers that had been financing the preview while awaiting general availability. Meanwhile, Santander Spain, which maintains primary infrastructure in Madrid, was one of the partners mentioned by Microsoft for early access in the West Europe region.


The Blind Spot


Microsoft and Anthropic avoid a public calculation this week: how much of the migration work from Bedrock to Foundry is already underway. Anthropic finished the last quarter with an annualized revenue in the neighborhood of $30 billion, more than three times what it had at the end of 2025. The leap did not come solely from the API; it resulted from Foundry and Bedrock seats converting from pilot to production, with over a thousand corporate clients paying more than $1 million per year on an annualized basis.


Every gigawatt of GB300 capacity that comes online now is capacity that comes out of the available batch for the next entrant, and the nod to European data zones signals that Anthropic does not see a risk of cannibalizing margins by spreading its footprint. The lingering question is whether the GA on Foundry pressures Bedrock to lower prices or whether AWS will respond with its own infrastructure. There are no public signals in that direction. What is clear is that Foundry is now in production on Blackwell Ultra.

Lead Analysis