DeepSeek releases V4 AI model, claims parity with OpenAI and Google
V4-pro carries 1.6 trillion parameters; a Huawei-adapted variant signals workaround to US chip controls
Briefing
- •DeepSeek has released a preview of its V4 large language model in two variants: V4-pro at 1.6 trillion parameters and V4-flash at 284 billion parameters, both with a 1 million token context window.
- •The Hangzhou startup claims V4 is competitive with leading closed-source models from OpenAI and Google DeepMind, and describes its efficiency as 'world-leading'.
- •Both models are open-source, continuing DeepSeek's strategy of releasing weights publicly, which amplifies competitive pressure on US frontier labs.
- •A version of V4 adapted for Huawei chip technology has been previewed, indicating DeepSeek is engineering around Nvidia supply constraints imposed by US export controls.
- •The release arrives roughly one year after DeepSeek's R1 model rattled global technology markets by demonstrating high performance at low reported training cost.
Analysis
- →DeepSeek V4's Huawei-adapted variant is the most consequential detail for markets. It demonstrates that China's AI labs are actively engineering around US export controls on Nvidia H100/H800 chips, which undermines the strategic premise that chip restrictions would degrade Chinese AI capability at a pace that preserves US frontier lab advantages. If Huawei silicon is now a viable substitute substrate, the export control regime loses its primary transmission mechanism, removing a structural tailwind for Nvidia's monopoly pricing in restricted markets.
- →DeepSeek V4's open-source release at frontier-parity performance compresses the implied value of closed-source model APIs from OpenAI and Google. Enterprise buyers now have a credible free alternative at 1.6 trillion parameters. This accelerates the commoditization of model inference and puts direct pressure on Microsoft Azure OpenAI Service revenue and Google Cloud AI API pricing, both of which have been a key incremental growth driver cited in recent earnings. The open-weight strategy also feeds directly into the enterprise software budget reallocation already documented in the ServiceNow and IBM earnings misses.
- →The one-year pattern from DeepSeek R1 to V4 shows a consistent cadence of releases that reset competitive benchmarks faster than US labs can widen their moat through capital expenditure alone. Microsoft's $140bn and Meta's $135bn AI capex commitments, paired with large-scale headcount reductions, are predicated on closed-model monetization holding at current margins. Each open-source DeepSeek release erodes that assumption, creating a scenario where the capex cycle delivers infrastructure with falling marginal revenue per model call, compressing the return on invested capital justification for the current spending wave.
Historical Context
- Jan 2025
DeepSeek R1's release triggered a single-day Nvidia selloff exceeding $500bn in market cap as markets priced the implication that frontier AI performance could be achieved at a fraction of assumed training cost. V4 arrives with the same open-source strategy and a Huawei hardware variant, compounding rather than repeating that event.
- Oct 2022
The Biden administration imposed sweeping export controls on advanced semiconductors to China, targeting Nvidia A100 and H100 chips. The policy was premised on hardware scarcity constraining Chinese AI development. DeepSeek's Huawei-adapted V4 is the first public evidence that a frontier Chinese lab has credibly engineered around that constraint.
- 2000–2003
The post-dot-com period showed that infrastructure buildout cycles can persist even as the revenue models funding them erode. Fiber was laid long after bandwidth pricing collapsed. The current dynamic where hyperscaler capex accelerates as open-source models commoditize closed-model pricing carries a structurally similar ROIC risk.
Connected Events
ServiceNow plunges 16% as IBM and NOW results deepen AI disruption fears across software sector
ServiceNow's 16% plunge on AI displacement fears, with IBM software revenue missing alongside it, confirms enterprise IT budgets are already reallocating away from incumbent SaaS toward AI infrastructure. DeepSeek V4's open-source release at frontier parity accelerates this reallocation by lowering the cost floor for capable AI, reducing the stickiness of closed-model contract commitments.
Microsoft launches first-ever voluntary buyout targeting up to 7% of US workforce
Microsoft's voluntary buyout of up to 7% of its US workforce, framed as capital redeployment into $140bn AI infrastructure spend, is directly exposed to DeepSeek V4: if open-source frontier models compress closed-API pricing, the revenue thesis that justifies converting payroll into capex weakens before the workforce reduction is complete.
SK Hynix posts record Q1 net profit on AI memory demand
SK Hynix's record Q1 profit arriving in line rather than ahead of estimates indicated AI memory demand is fully priced into consensus. DeepSeek V4's efficiency claims, if validated, reduce per-inference memory bandwidth requirements, introducing a downside scenario for HBM demand volumes that the current SK Hynix valuation does not price.
%3Amax_bytes(150000)%3Astrip_icc()%2FGettyImages-2271950132-7241f823699b42b4b10d142442f2680d.jpg&w=3840&q=75)