CoreWeave Breaks MLPerf Record With DeepSeek-V3 Training
8h ago|4 min readStandard
Fazen Markets Editorial Desk
Collective editorial team · methodology
coreweavemlperf
Fazen Markets Editorial Desk
Collective editorial team · methodology
Trades XAUUSD 24/5 on autopilot. Verified Myfxbook performance. Free forever.
Risk warning: CFDs are complex instruments and come with a high risk of losing money rapidly due to leverage. The majority of retail investor accounts lose money when trading CFDs. Vortex HFT is informational software — not investment advice. Past performance does not guarantee future results.
Specialized cloud provider CoreWeave achieved the fastest recorded time for training the DeepSeek-V3 large language model in industry-standard MLPerf benchmarks, Investing.com reported on 16 June 2026. The company completed the benchmark training run in 11.2 minutes, setting a new performance record. This result comes amid a heightened period of competitive benchmarking among AI infrastructure vendors and follows a record $8.5 billion investment round into CoreWeave in May 2026.
The MLPerf consortium has become the definitive arena for proving the performance of AI hardware and software stacks since its inception in 2018. CoreWeave's victory follows a series of record-setting runs by other hyperscale and specialized providers. In April 2025, Google's TPU v5 platform set a then-record of 16.5 minutes for a similar benchmark. The current macro backdrop for AI infrastructure features intense capital expenditure. The 10-year Treasury yield stands at 4.31%, influencing the cost of capital for the data center buildouts underpinning this competition.
Several factors converged to trigger this specific event. The release of the DeepSeek-V3 model architecture provided a fresh, demanding standard for benchmarking. Concurrently, CoreWeave secured a massive $8.5 billion capital infusion, allowing it to deploy and optimize its latest generation of NVIDIA hardware, specifically the H200 platform. This capital advantage translated directly into an ability to scale its cluster size and refine its software stack for the June 2026 submission window, aiming to demonstrate a tangible return on its funding.
The catalyst chain is straightforward. Massive investment enables procurement of cutting-edge hardware. That hardware, combined with proprietary software optimization, yields a measurable performance lead in a recognized industry test. This lead is then marketed to enterprise clients and developers as proof of superior efficiency, directly impacting customer acquisition and market share in a sector where performance translates to lower training costs and faster iteration cycles.
CoreWeave's submission posted a training time of 11.2 minutes for the DeepSeek-V3 model. This represents a 32% speed improvement over the previous record of 16.5 minutes set 14 months prior. The benchmark run utilized a cluster of 4,096 NVIDIA H200 GPUs. The performance translates to a sustained aggregate compute throughput exceeding 65 exaflops.
The table below contrasts CoreWeave's latest result with a key prior submission:
| Metric | CoreWeave (Jun 2026) | Google TPU v5 (Apr 2025) | Change |
|---|---|---|---|
| Training Time | 11.2 minutes | 16.5 minutes | -32% |
| Hardware | 4,096 NVIDIA H200 | Undisclosed TPU v5 Pod | N/A |
| Model | DeepSeek-V3 | Comparable LLM | N/A |
This performance gain significantly outpaces the broader market's efficiency improvements. The S&P 500 Information Technology Index has gained 12% year-to-date, reflecting general optimism, but CoreWeave's 32% benchmark leap is a discrete, outsized technical advance. The result also underscores the continued dominance of NVIDIA's GPU architecture in raw performance benchmarks, despite rising competition from custom silicon alternatives.
Second-order effects will manifest across several sectors. Direct beneficiaries include NVIDIA (NVDA), as record results reinforce demand for its flagship H200 and Blackwell GPUs. CoreWeave's private market valuation, last pegged at $55 billion, will face upward pressure, positively impacting its investors and the private equity sector focused on AI infrastructure. Companies reliant on large-scale AI model training, such as OpenAI, Anthropic, and Tesla (TSLA) for its autonomous driving models, benefit from lower costs and faster development cycles, potentially accelerating their own product roadmaps.
Conversely, the result poses a challenge for hyperscale cloud providers like Amazon Web Services (AMZN), Microsoft Azure (MSFT), and Google Cloud (GOOGL). While they compete on integrated suites of services, a pure performance gap in core training workloads could influence cost-sensitive enterprise AI projects. The semiconductor sector sees a mixed impact; while NVIDIA gains, competitors like AMD (AMD) and Intel (INTC) must demonstrate comparable or superior performance in future benchmarks to maintain credibility in the AI accelerator space.
A key limitation of this analysis is that MLPerf benchmarks measure peak performance on a controlled workload, not necessarily real-world, sustained cost-efficiency over a multi-month training campaign. Network bottlenecks, software reliability, and actual power consumption are critical factors not fully captured in the headline 11.2-minute figure. Positioning data shows institutional flow moving towards the AI infrastructure layer. Hedge funds are increasingly long NVIDIA and exploring pre-IPO secondary market positions in CoreWeave, while short interest has risen in traditional data center REITs perceived as less agile.
Market participants should monitor several imminent catalysts. NVIDIA's Q2 2026 earnings report on 20 August will provide insight into H200 and Blackwell rollout velocity and demand signals from large cloud providers. The next MLPerf submission window in September 2026 will reveal if competitors have closed the performance gap. The potential CoreWeave initial public offering, rumored for the first half of 2027, will be a major liquidity event for private investors and a fresh public comp for the sector.
Key technical levels to watch include NVIDIA's stock price relative to its 50-day moving average, a breach of which could signal a sentiment shift. In the bond market, watch the yield on the iShares Semiconductor ETF (SOXX) for signs of changing risk appetite towards the capital-intensive chip sector. Monitoring the valuation multiples of public cloud companies versus their revenue growth from AI services will indicate whether the market is pricing in sustained infrastructure advantage or a more commoditized outcome.
Faster training time directly reduces the largest variable cost in AI development: cloud compute expenses. A 32% speed improvement, as demonstrated by CoreWeave, can translate to a proportional reduction in the cost to train a frontier model, which often runs into tens of millions of dollars. This efficiency gain allows research teams to iterate more quickly, testing more model architectures and datasets within a fixed budget. It also lowers the barrier to entry for smaller firms aiming to develop competitive models, potentially increasing innovation and competition in the AI application layer.
Vortex HFT is our free MT4/MT5 Expert Advisor. Verified Myfxbook performance. No subscription. No fees. Trades 24/5.
Position yourself for the macro moves discussed above
Start TradingSponsored
Open a demo account in 30 seconds. No deposit required.
CFDs are complex instruments and come with a high risk of losing money rapidly due to leverage. You should consider whether you understand how CFDs work and whether you can afford to take the high risk of losing your money.