Fireworks contracts 5MW of Nvidia Blackwell Ultra capacity at Era4 data center in UK - Data Center Dynamics

Fireworks locking 5MW of Blackwell Ultra at Era4 isn't a procurement deal—it's a positioning play in the inference cloud war. Five megawatts translates to roughly 40 GB200 NVL72-class racks. That's production-scale inference, not R&D. The real chain reaction sits upstream. Blackwell Ultra's HBM4 bandwidth and expanded NVLink domain will cut per-token inference costs by 30-40%, directly compressing margins for Together AI, Replicate, and similar platforms. The binding constraint is no longer chip design—it's TSMC CoWoS-L packaging capacity and HBM4 supply from SK Hynix. Power, not silicon, is becoming the industry's scarcest resource. On compliance, Era4's status as a UK sovereign data center signals Fireworks is hedging against EU AI Act deployment-location requirements while avoiding cross-border data exposure. But 5MW of load on the UK grid is non-trivial; 2025 data-center power quotas are tightening, and marginal expansion costs will spike. The competitive window is roughly 12 months. AMD MI500 and Google TPU v6 will erode Blackwell Ultra's performance premium. Fireworks must convert this hardware advantage into locked customer contracts before the premium dilutes. The 12-24 month thesis: inference will fully overtake training as the dominant compute consumer. Whoever secures the cheapest power-plus-latest-silicon pairing first owns structural pricing power. Fireworks is betting on exactly that.