DeepSeek V4's Million-Token Window Is Real Progress — But It Doesn't Prove What Enterprise Adopters Think It Does

Here's the distinction that matters: our forecast isn't asking whether million-token windows exist — they do, in multiple models. It's asking whether Fortune 500 companies are running production workflows on them at scale, with acceptable latency, cost, and reliability. Those are entirely different questions, and today's DeepSeek announcement addresses only the first. The V4 release is strong circumstantial evidence that the technology is maturing and commoditizing — an open-source 1M-token model from a non-US lab that's simultaneously seeking $300M+ in external funding represents meaningful capability diffusion. But capability diffusion is not enterprise production deployment.

The gap in our model — and we want to name this directly — is that we don't have good visibility into what Fortune 500 companies are actually running versus piloting. The 45% reflects our judgment that latency and cost remain prohibitive for most enterprise workflows at million-token scale, and that RAG architectures are cheaper and more reliable for the document retrieval use cases where long context would otherwise shine. But if major cloud providers start shipping managed million-token APIs with enterprise SLAs at competitive price points — which DeepSeek's open-source release could accelerate by forcing Google and Amazon to compete on price — that cost-reliability calculus shifts faster than our current model assumes.

The counterevidence we're taking seriously: DeepSeek's decision to push 1M tokens as a flagship feature, combined with its open-source release strategy, suggests the lab believes this is where enterprise demand is heading. They're not building for hobbyists. Tencent and Alibaba circling for investment at a $20B+ valuation implies the Chinese tech establishment agrees. That's market signal, not just technical achievement. It's also worth noting that the use case DeepSeek specifically highlights — sending entire codebases as a single prompt — is exactly the kind of task where enterprises running Copilot-style coding tools would benefit most. That's a narrower but more plausible production use case than 'entire document corpus ingestion.'

What would move us to 60%+? Evidence that a named Fortune 500 company has migrated a production workflow — not a pilot, not a proof of concept — from RAG to million-token context, with measurable throughput and cost metrics. What would drop us to 30%? Q2 enterprise software earnings calls showing AI infrastructure spend growing but with explicit callouts that long-context features remain in evaluation rather than production. We're watching Salesforce, ServiceNow, and Microsoft's enterprise Azure AI segment disclosures most closely for this signal.