On October 23, 2026, NVIDIA and TCV Fund invested in the expansion of the Series B round for Kazakh AI startup Higgsfield—an image and video generation platform for brands and studios that raised $400 million in August 2026 at a $5.4 billion valuation and announced in September that it had exceeded $1 billion in annualized revenue. The announcement coincided with Jensen Huang's first visit to Astana, where NVIDIA confirmed that the country's largest telecom, Kazakhtelecom, has already deployed 880 GPUs. In parallel, the Firebird project received a US license to export 45,000 Grace Blackwell accelerators for the first phase of an AI factory in Ekibastuz, with a commercial launch scheduled for the third quarter of 2027.

image
image

What happened

NVIDIA and TCV Fund invested in the expansion of Higgsfield's Series B round, which opened in August 2026 with a $400 million raise at a $5.4 billion valuation; in September, the company reported surpassing $1 billion in annualized revenue. The investment was announced during Jensen Huang's first visit to Astana: there, NVIDIA confirmed that Kazakhtelecom has deployed 880 GPUs in Astana and Almaty—400 H200s and 480 B300s. In parallel, Firebird received a US license to export 45,000 NVIDIA Grace Blackwell (GB300) accelerators for the first phase of an AI factory in Data Center Valley in Ekibastuz. The first phase is estimated at approximately $5 billion, with a stated capacity reaching 109 MW, and a commercial launch planned for the third quarter of 2027; Kazakhtelecom is the project partner.

Context

Understanding the deal is aided by the structure of the players. Higgsfield is a vertical image and video generation platform for brands and studios, with more than 300 engineers working on-site in Kazakhstan. Kazakhtelecom, the country's largest telecom, acts as the operator of the computing infrastructure and partner for the Firebird factory in Data Center Valley. Kazakhtelecom's initial fleet is heterogeneous: it mixes GPUs from two generations, Hopper and Blackwell, which complicates the software stack—CUDA and driver version compatibility, task scheduling, and network topology—and immediately tests the competencies of the hub's future operations team. The second tier of infrastructure is built on Grace Blackwell GB300 superchips: for this class of hardware, the bottlenecks are power, cooling, and the network fabric, not the availability of chips. The national expectation sets the frame: Kazakhstan declared 2026 the Year of Digitalization and AI and positions itself as a computing hub for Europe.

Why this matters for the industry

For the industry, Kazakhstan demonstrates an inverse model of AI infrastructure development: first, an engineering team and a working product with revenue were established in the country, and GPUs and capital followed the talent. The stated volumes add up to one of the largest GPU clusters in the region: 45,000 Grace Blackwells for Firebird plus Kazakhtelecom's plans to add more than 1,000 GPUs and more than 1,000 Vera Rubin accelerators in the Ekibastuz region, which could reshape the map of sovereign computing for Europe and Asia. Builders need to distinguish between what is working and what is announced: the real environment for inference, LoRA, and RAG benchmark pilots is the deployed fleet of H200s and B300s with operator access, while the gap between the deployed and announced fleets exceeds fifty-fold, making it premature to plan production for GB300s before supply is confirmed. The hub's fate will ultimately be determined not by the number of chips, but by the quality of operations: uptime, latency, price, and portability between accelerator generations. In parallel, teams should anticipate growing competition for regional engineers.

Why this matters for users

For readers and teams looking at the region, what is available today is not computing capacity, but early signals: sources do not mention a new public API or product that can be connected right now, and access terms to the working cluster—tariffs, SLAs, and utilization statistics—are not disclosed, so a reasonable first step is to request these parameters directly from the operator. If the schedule holds, by the third quarter of 2027, the first phase of the factory could handle regional inference workloads and become a hosting option for European and Asian teams; simultaneously, expect the first commercial offers for access and rental from Kazakhtelecom and Firebird. Benchmarks to watch: GB300 deliveries to Ekibastuz, the site's energy readiness for 109 MW, actual utilization of already working capacities, and the first Vera Rubin deliveries by 2028. Higgsfield provides a separate storyline: the Central Asian startup became Kazakhstan's first AI unicorn, taking about 18 months to reach $1 billion in annualized revenue.

What is currently unknown / limitations

Key figures have not yet been independently verified: the $1 billion annualized revenue and the status of the country's first AI unicorn come from Higgsfield's official materials, and the investment amount from NVIDIA and TCV is not disclosed in publications. It does not follow from the sources that the company is developing its own models: there are no scientific publications or benchmarks in the corpus, so the news speaks to the product and revenue, not to a new methodology or a new quality of models. The 45,000 GB300s remain a supply obligation for the third quarter of 2027: there is currently no data on actually delivered batches or the site's energy readiness for 109 MW, and a lack of physical progress on hardware by spring 2027 would increase the probability of a schedule shift.

Sources

Author

Look at AI, editorial team