DeepSeek Building Custom AI Chip to Break Free From Nvidia

DeepSeek's $7B raise and chip ambitions signal vertical integration race in AI; pricing power demonstrates efficiency now trumps capability

DeepSeek closed its first institutional funding round at $7B with founder Liang Wenfeng retaining complete control while funding 40% himself, simultaneously developing proprietary inference chips to reduce reliance on Nvidia and Huawei. The chip effort, which began roughly a year ago, targets the fastest-growing segment of AI computing demand and includes increased hiring of chip-design engineers, though success remains uncertain given manufacturing constraints and U.S. export controls on advanced foundries. This mirrors OpenAI's Jalapeno and Anthropic's chip exploration, signaling a broader shift toward vertical integration across models, infrastructure, and silicon. DeepSeek's pricing advantage underscores why this integration matters: V4 Pro costs $0.87 per million tokens versus Claude Fable 5's $50—a 57x difference that demonstrates capability alone no longer determines winners. As users and agents optimize for cost-performance tradeoffs week-to-week, efficiency and vertical control of the stack will constantly reshuffle which models capture value, fundamentally changing how VCs should evaluate AI companies.

SpaceX S1 filing reveals AI infrastructure economics: Anthropic paying $5.30/GPU-hour for 325,000 units ($1.25B/month), Google $11.50/hour for 110,000 units

SpaceX's S-1 registration statement discloses concrete GPU procurement terms with major AI labs, showing Anthropic's deal alone implies a 21-month payback on SpaceX's $26.6B capex investment. The contracts include unusually permissive 90-day cancellation clauses, signaling both the scale of AI infrastructure demand and the risk that customers retain flexibility to shift capacity allocation. Anthropic's lower per-unit cost ($5.30 vs. Google's $11.50/hour) reflects volume leverage and negotiating power, with the company locking in 325,000 units versus Google's 110,000. Concurrently, memory supply is emerging as the tighter constraint than GPUs themselves. Micron projects 66% operating margins in 2026 on AI and HBM demand, while SK Hynix raised $26.5B on Nasdaq and China's CXMT remains years behind in HBM production, creating a durable oligopoly dynamic. If long-term contracting persists, this could break the semiconductor industry's 40-year commodity cycle and lock in sustained infrastructure spending.

Spectral Compute raises $6M to break Nvidia's CUDA lock-in; SCALE compiler enables instant CUDA-to-AMD portability

Spectral Compute, a British software startup, raised $6M in seed funding to solve the "Nvidia tax" by enabling CUDA code to run natively on non-Nvidia GPUs without costly rewrites. Its SCALE toolkit compiles CUDA programs for AMD GPUs in seconds, compared to months of manual porting required by existing tools, giving enterprises genuine multivendor procurement strategies. While current support is limited to AMD, the company plans to expand to all AI accelerators. The round was led by Costanoa with participation from Crucible, signaling investor conviction that CUDA portability is a critical infrastructure bottleneck as GPU demand outpaces Nvidia supply.

DeepSeek Developing Its Own AI Inference Chip to Reduce Nvidia and Huawei Reliance

Reuters reported that Chinese AI startup DeepSeek is developing its own custom AI chip focused on inference workloads, marking a major strategic shift for the company. Main takeaways:

• DeepSeek's chip targets inference (not training), the fastest-growing segment of AI compute demand, positioning it to reduce dependence on Nvidia and Huawei amid U.S. export controls
• The company has increased hiring of chip-design engineers over recent months and is in early-stage discussions with chip-design, foundry, and memory companies
• DeepSeek's effort follows OpenAI's Jalapeno announcement and reflects broader industry trend of AI labs building proprietary inference hardware to control costs and reduce reliance on Nvidia
• Manufacturing remains a critical constraint: U.S. bans prevent Chinese designers from accessing advanced foundries, and separate curbs limit access to high-bandwidth memory essential for AI inference chips
• The chip push coincides with DeepSeek's first external funding round ($7B at $52-59B valuation), reversing years of rejecting outside capital