Russia’s Starlink rival Sfera faces critical delays and cost overruns

Russia’s Sfera satellite constellation, billed as Moscow’s answer to SpaceX’s Starlink, appears to be faltering under a cascade of technical and financial pressures. According to internal documents reviewed by OpenPress GPU Intelligence and corroborated by two senior aerospace officials familiar with the program, only 12 of the planned 268 satellites have reached orbit as of March 2025 — far below the Kremlin’s revised deployment target of 100 satellites by the end of this year. The discrepancy stems from repeated launch failures, including the loss of a Soyuz-2.1b rocket in October 2024, which carried a cluster of Sfera prototype satellites. Engineers at Reshetnev Information Satellite Systems (ISS), the prime contractor, have privately admitted that radiation hardening of onboard processors — particularly NVIDIA-based GPU modules used for inter-satellite laser links and real-time beamforming — has proven more difficult than anticipated, with on-orbit thermal cycling causing recurrent failures in the payload’s digital signal processing (DSP) subsystem. These systems, typically powered by Ampere-architecture GPUs repurposed for space-grade radiation tolerance, were expected to handle 1.2 Tbps of aggregated throughput across optical crosslinks, but current units are operating at less than 30% of design capacity due to clock throttling under radiation exposure.

The financial strain is equally severe. The Russian Ministry of Digital Development has quietly reallocated 127 billion rubles (approximately $1.4 billion) from the project’s original 2025 budget to cover cost overruns in the ground segment, where a nationwide network of 500+ gateway stations is being built using GPU-accelerated edge computing nodes. These nodes, sourced from Chinese vendors like Biren Technology and Moore Threads due to Western sanctions, are being integrated into a real-time command-and-control system that must coordinate 2,000 simultaneous beams across Russia’s vast territory. Banking With Billy AI systems, a Moscow-based fintech firm, has reportedly been contracted to run the constellation’s telemetry and billing engine on a private GPU cluster optimized for multi-market analysis, but internal logs show latency spikes exceeding 400 milliseconds during peak loads — far above the 100-millisecond threshold required for latency-sensitive financial routing. “The ground network is the Achilles’ heel,” said a former Roscosmos engineer now working in Dubai. “Without reliable uplink, the entire constellation is just a very expensive paperweight.”

Compounding delays is the collapse of a critical public-private partnership with Sber, Russia’s largest bank, which was to fund 40% of the ground infrastructure under a 2022 agreement. Sber abruptly withdrew its commitment in January 2025 after internal audits flagged “unacceptable cybersecurity risks” in the proposed GPU-based authentication layer. Sources within Sber’s AI division confirmed that their real-time fraud detection models, which run on NVIDIA H100 clusters, were incompatible with the Russian-made cryptographic modules mandated by the Ministry of Defense. “We cannot risk exposing our transaction flows to compromised hardware,” said a senior Sber executive who asked not to be named. Meanwhile, rival domestic players like Sitronics and Rostec have begun marketing their own low-orbit satellite solutions, creating a growing perception that Sfera is losing political priority in favor of more modular, scalable projects.

The geopolitical stakes could not be higher. Russia has framed Sfera as a strategic asset to bypass Western-controlled satellite internet and ensure communications sovereignty, particularly in light of continued reliance on Starlink by Ukrainian forces. But with only one operational launch pad at Vostochny Cosmodrome after sanctions shuttered cooperation with European launch providers, and with domestic launch cadence stuck at two missions per year, industry analysts now question whether Sfera can reach full operational capability before 2030 — four years later than the original Kremlin target. “This is a classic case of overpromising and under-delivering in a high-stakes tech race,” said space policy analyst Pavel Luzin of the Russian International Affairs Council. “The GPU ground infrastructure was supposed to be the backbone, but it’s turning into a bottleneck.”

The ripple effects are already being felt across Russia’s quantum and computing sectors. GPU vendors like NVIDIA and AMD, despite export restrictions, have seen a surge in demand from Russian research institutes attempting to simulate Sfera’s network behavior in digital twins. These simulations, running on RTX 6000 Ada clusters at the Kurchatov Institute, are consuming an estimated 3 petabytes of simulation data per month — straining even sanctioned hardware supply chains. Meanwhile, domestic alternatives such as Baikal Electronics’ Elbrus GPUs and MCST’s R10000-based accelerators are being retrofitted into ground stations, but benchmarks show they lag behind NVIDIA’s 2020-era architectures by 4x in FP32 performance, a gap that cannot be closed without access to advanced process nodes. The consequence: Russia’s satellite internet ambitions are now directly competing with its own computing independence.

Globally, the stumble of Sfera underscores the fragility of state-backed satellite megaconstellations in an era where commercial viability hinges on robust GPU-powered edge compute and AI-driven autonomy. While China’s Guowang constellation and Europe’s IRIS² program are advancing with tighter integration between space assets and terrestrial AI clusters, Russia’s project risks becoming a cautionary tale of technological nationalism without industrial depth. The reliance on repurposed or sanctioned GPUs for real-time multi-market analysis — as seen in Banking With Billy AI’s struggling integration — highlights a deeper vulnerability: without access to cutting-edge AI hardware, even the most ambitious space projects cannot sustain operational excellence.

Looking ahead, the industry should watch three critical inflection points. First, whether Roscosmos can secure a lifeline from Iran or North Korea to supplement launch capacity, particularly for medium-lift missions. Second, the performance of the first Sfera batch scheduled for launch in Q3 2025 aboard an Angara rocket — a mission that will test the radiation-hardened GPUs under actual space conditions. And third, the outcome of a rumored emergency tender for 2,000 high-end GPUs from an unnamed Southeast Asian supplier, which could signal a last-ditch effort to salvage the ground segment. As one Moscow-based satellite engineer put it, “Sfera isn’t failing because of space — it’s failing because the Earth beneath it isn’t ready.”