AI Intelligence

DeepSeek Nears $7.4 Billion as US Firms Turn East

Vijay Bhagwati

04 Jun 2026 • 5 min read

US companies are paying Chinese AI labs directly as cost pressure overrides geopolitical caution.

June 4, 2026 | Reading time: 9 minutes | Issue #179

DeepSeek is finalizing its first external fundraising round at a valuation approaching $60 billion, according to a June 3 report from the South China Morning Post. The Hangzhou-based lab has secured over $7.4 billion in commitments from Tencent, NetEase, JD.com, CATL, and several venture firms, with founder Liang Wenfeng contributing roughly 20 billion yuan himself. The round marks a six-fold jump from DeepSeek's $10 billion valuation in April and ends the company's long refusal of outside capital.

On the same day, DeepSeek topped the June trending-software index published by New York-based Ramp, which tracks first-time corporate purchases from software vendors. According to Ramp economist Ara Kharazian, US firms are sending data directly to DeepSeek's servers rather than hosting its open-source models internally. The migration suggests cost pressure is overriding geopolitical caution: DeepSeek's flagship V4 Pro is priced at roughly one-quarter its US equivalents following a permanent 75% reduction announced in May, and benchmark firm Artificial Analysis has ranked it among the top models globally on an intelligence-per-dollar basis.

The convergence of a record Chinese fundraise and first-time US enterprise adoption signals a structural shift in the model market. DeepSeek is no longer a viral GitHub repository or a cautionary tale about Chinese AGI. It is a commercial competitor that American companies are paying directly. The question is whether Washington treats that as a trade issue or a security one.

Companies' AI bills are bigger than ever — and coming due

Bloomberg published an analysis on June 4 warning that corporate AI bills are reaching unsustainable levels and that the next industry test is proving ROI. The piece notes that token spend has become one of the fastest-growing cost centers in engineering organizations, with many teams unable to attribute AI usage to concrete outcomes. The warning arrives as Google, Anthropic, and SpaceX have collectively raised more than $220 billion in one week — capital that is now hitting balance sheets rather than venture pipelines.

OpenAI expands Codex to non-developers

OpenAI announced on June 2 that Codex is expanding beyond software developers with role-specific plugins, inline annotations, and a preview of Sites for sharing interactive apps. The company disclosed that non-developers — analysts, marketers, bankers, and researchers — now account for 20% of Codex users and are growing three times faster than the developer base. The move frames Codex as a general-purpose knowledge-work platform, not merely an IDE extension.

Mistral enters Physics AI

Mistral AI introduced a new research stream called "Physics AI" on May 27, with models designed to predict the behavior of physical systems across engineering and materials science. The announcement, coupled with the acquisition of Paris-based simulation startup Emmi, places Mistral in a category distinct from chat model competition: industrial physics simulation. The move adds a European counterweight to OpenAI's materials-science push and NVIDIA's Omniverse physics engine.

OpenAI ships Rosalind update

OpenAI updated its GPT-Rosalind life-science series on June 3 with stronger medicinal-chemistry and genomics capabilities, along with a new benchmark called LifeSciBench that evaluates end-to-end scientific workflows across evidence handling, design, and wet-lab troubleshooting. The model is now available to trusted enterprise partners and select government biodefense teams under a restricted-access deployment structure.

Open-Source Pulse

The Ramp index for June is not just a one-off purchase blip. It reflects a broader migration from proprietary API spending to open-weight inference platforms. DeepSeek, Fireworks AI, Fal AI, and DeepInfra all ranked among the month's trending vendors, suggesting that open-source capabilities have reached parity with premium closed models on legal and enterprise benchmarks.

Fireworks AI disclosed this week that Zhipu AI's GLM 5.1 ranked highest among open-source models on Harvey's Legal Agent Benchmark, trailing only Anthropic's Claude Opus 4.7 and matching OpenAI's GPT-5.5. DeepSeek's V4 Pro and Moonshot's Kimi K2.6 also placed within the viable band. The pattern is unmistakable: the gap between open and closed is closing on commercial tasks, and the price differential is forcing procurement officers to reconsider API lock-in.

Builder's Corner

OpenAI's Codex expansion on June 2 includes three builder-facing features: role-specific plugins for non-developers, inline annotations for refining output without leaving the canvas, and a Sites preview for turning Codex sessions into shareable interactive apps. Inside OpenAI, non-technical teams reportedly use Codex to build internal apps, executive materials, and dashboards. At Zapier, Codex pulls knowledge from Slack, Google Docs, and Coda to generate postmortems and feature tickets. At NVIDIA, researchers use it for machine-learning infrastructure scripting.

The through-line is that agentic coding is moving downstream from professional developers to knowledge workers who need custom tools but lack frontend skills. For builders, the implication is that the market is shifting from code generation to workflow generation — from writing functions to orchestrating human-AI hybrid pipelines.

India Lens

India signed an agreement in May with the Abu Dhabi-backed G42 group and US chipmaker Cerebras to deploy 64 AI supercomputers inside India, according to a June 1 report from Rest of World. The deal gives India a sovereign compute pathway outside the $45 billion of existing commitments it has with Amazon, Microsoft, and Google. Data will remain under Indian governance rules, but G42 declined to disclose ownership terms for the hardware after installation.

The Cerebras choice is notable. While India's national AI program runs on Nvidia processors, Cerebras offers inference-optimized wafer-scale chips rather than training accelerators. The distinction signals that India's priority is deploying AI across agriculture, healthcare, and public services rather than training frontier models from scratch. This is a pragmatic sovereignty play: own the compute, rent the model.

The View

DeepSeek's fundraise and OpenAI's Codex expansion are two sides of the same pressure: cost. DeepSeek is winning on price. OpenAI is winning on distribution. Both strategies assume that frontier models will be commodities. DeepSeek is betting that intelligence-per-dollar is the purchasing criterion, which works until geopolitical sanctions cut off server access. OpenAI is betting that embedding agents into every knowledge-worker workflow creates switching costs that survive any price war, which works until open-source alternatives can replicate the orchestration layer.

The middle ground is where the money will be made: the companies that neither build models nor host them, but that bolt models into the enterprise stack. Indian IT firms, European integrators, and American low-code platforms are all positioning for that middle layer. The model war is becoming an integration war, and the winners will be the ones who can prove the bill is worth paying.

The Miss

A blog post published on June 4 by Pensero argued that most engineering organizations cannot explain what they are getting in return for token consumption. Token spend becomes the new cloud bill: growing fast, poorly attributed, and only questioned when it becomes uncomfortable. The piece received modest attention on Hacker News but frames a structural risk: if the next round of budget scrutiny asks AI teams to show ROI, many will not have the data. The labs that survive a downturn will be the ones that made their models measurable before finance asked.

Pull Quotes

"In probably the biggest sign that companies are looking for cheaper alternatives to OpenAI and Anthropic, some are willing to use cheaper, Chinese models, sending US data back and forth from China-hosted servers." — Ara Kharazian, Ramp Economics Lab

"The real question in enterprise AI is not who builds the most capable model. It is, 'Who can make AI work inside messy, complex enterprise environments that have accumulated decades of process debt, data debt, technology debt, and cultural debt?'" — Ashwin Venkatesan, HFS Research

"Token winter is coming." — Pensero, Hacker News

Out

The next frontier is not model intelligence. It is procurement intelligence.

By Neo