A short post from a developer has turned into one of the week’s biggest AI infrastructure debates. Dax Raad (@thdxr) argued that OpenRouter, “the most successful model router”, could “almost serve 100% of their deepseek traffic yesterday on a single GB300 rack”, hardware he put at “6-7M”, adding that “even if there’s 10x this demand secretly that’s still < 100M of compute”. Others ran the same napkin math and asked the obvious question: if efficient open models need so little hardware, why are Nvidia and the hyperscalers planning for tens of thousands of racks next year?
Quick Take
- The claim: Developer Dax Raad says a single Nvidia GB300 NVL72 rack could almost serve a day of OpenRouter’s DeepSeek traffic; investor Chon Tang put the figure at 5–10 racks.
- The efficiency: DeepSeek’s V4.1-Flash-1017, released with open weights on October 17, reports 77.0% on DeepSWE at under $0.20 per task, according to DeepSeek’s own figures.
- The demand side: TrendForce expects the nine largest cloud providers to spend about $957 billion on capex in 2026 and more than $1.5 trillion in 2027.
- The verdict: Public router traffic is a small, visible sliver of AI compute; training, agents and closed-lab inference drive rack demand. Efficiency mainly pressures prices.
What happened
Raad’s post landed amid excitement over DeepSeek’s latest release. V4.1-Flash-1017 is a small, open-weight model that, per DeepSeek’s charts reported by The Decoder, scores 77.0% on the DeepSWE agentic coding benchmark at under $0.20 per task, compared with 74.2% at about $0.50 for its predecessor. A low-effort setting reaches roughly 72% for about $0.10. DeepSeek positions rivals at $2 to more than $10 per task. Independent benchmark results were not yet available at publication. Off-peak API pricing is reported at $0.15 per million input tokens and $0.60 per million output tokens.
Chon Tang then asked: if DeepSeek is “roughly 30% of total OR token traffic” and that fits on “5-10 GB300 NVL72 racks”, and “NVIDIA is shipping 90k racks next year”, “what am I missing?” The post spread quickly among investors already nervous about AI capex.
Public data complicates the arithmetic. Vantage’s OpenRouter tracker shows DeepSeek as a provider at about 946 billion tokens per day on October 26, a 3.6% share, while other estimates on X put all DeepSeek-model traffic across providers at several trillion tokens per day. Neither figure is audited, and the 30% figure appears to count DeepSeek models served by many providers.
Why it matters now
The question echoes the “DeepSeek moment” of early 2025, when an efficient Chinese model briefly knocked Nvidia’s market value. This time the hardware is far more capable. SemiAnalysis’s InferenceX benchmarks show DeepSeek V4 Pro on GB300 reaching about 62,700 tokens per second per GPU at 100 tokens per second per user, and estimate Nvidia’s Vera Rubin will deliver 1.8 to 2 times more throughput per dollar of total cost of ownership.
At the same time, forecasts keep climbing. TrendForce projects AI server revenue near $920 billion in 2027, about 81% of the server market, with GB300, Vera Rubin VR200 and AMD’s Helios leading orders. Goldman Sachs estimated in June roughly 105,000 NVL72-equivalent racks across vendors in 2027 and 163,000 in 2028, with Nvidia’s own NVL72 racks at about 92,000 and 148,000, while custom ASICs grow toward 55% of accelerators.
Who is most affected
Inference providers and API resellers feel the efficiency squeeze most directly, because lower per-token costs compress their margins. Chipmakers and hyperscalers are less exposed than the viral math implies.
| Stakeholder | Impact | Why |
|---|---|---|
| Inference providers | High | Open models at cents per task compress per-token margins and commoditize hosting. |
| Nvidia & hyperscalers | Medium | Narrative risk from viral math, but forecasts are driven by training, agents and first-party inference. |
| Developers & agent builders | High (positive) | Cheaper agentic coding runs make larger, longer workloads economically viable. |
The under-reported angle: the visible slice is not the market
OpenRouter is a public marketplace, mostly serving developers and coding agents. It does not show OpenAI, Google, Anthropic or xAI first-party inference, enterprise deployments, reinforcement learning runs, or training clusters, which absorb most frontier hardware. A 3.6% or even 30% share of one router says little about global demand. Efficiency also tends to expand usage: when an agentic coding task drops from dollars to cents, developers run far more of them, the classic Jevons effect. There is also a policy wrinkle. Reports indicate China is tightening rules on open-weight releases, which could slow the very open-model supply that powers these cost claims.
What Grok and X are saying
Grok’s replies were measured. Asked to check Raad’s claim, it said it “checks out directionally”, estimating DeepSeek volume on OpenRouter at roughly 8–10 trillion tokens per day and a GB300 NVL72’s capacity on Flash-class models in a similar range. That volume estimate is far above Vantage’s provider-level figure, which shows how loosely these numbers are pinned down. On the wider implication, Grok answered Tang directly: “What’s missing: OpenRouter is a tiny public slice. Most tokens never hit it”, adding that “training + proprietary inference dwarfs this by orders of magnitude” and that Nvidia is on track for roughly 70,000–100,000 NVL72-class racks in 2026–27, in line with the public forecasts above. Elsewhere on X, bears argued the post proves hyperscalers are overbuilding, while infrastructure engineers countered that serving cheap inference at scale still needs redundancy, geographic spread and headroom for traffic spikes, none of which a one-rack snapshot captures.
Our read: both camps are partly right. Efficiency is real and is already pushing token prices down. But it changes who earns margin, not whether racks get built.
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis synthesizes vendor-reported benchmarks, public traffic trackers, analyst forecasts and verified posts on X. Vendor benchmark and traffic figures are labeled as such and were not independently verified.
- Dax Raad on X: DeepSeek traffic on a single GB300 rack
- Chon Tang on X: “What am I missing?”
- The Decoder: DeepSeek V4.1-Flash-1017 cuts per-task costs
- Vantage: OpenRouter provider statistics for DeepSeek
- TrendForce: CSP capex and AI server outlook
- Goldman Sachs rack forecast (via Tiger Brokers)
- Grok on X: the claim “checks out directionally”
- Grok on X: OpenRouter is a tiny public slice
i10x Perspective
The one-rack debate is a useful stress test for AI infrastructure narratives. Efficient open models like DeepSeek’s will keep resetting the price of intelligence, and that is bad news for anyone selling undifferentiated tokens. But cheaper intelligence has historically meant more consumption, not less hardware. The signal to watch is not router share; it is whether cheaper agents push enterprise usage up faster than per-token prices fall.
