DeepSeek launches 1M‑token V4 preview with lower prices, pressuring rivals
China’s DeepSeek ships two MoE model previews that it says approach frontier performance, while Cohere moves on a sovereign AI merger and Nvidia backs core data infrastructure.
One-Line Summary
Global AI competition tightens as DeepSeek ships a cheaper 1M‑token model preview, Cohere pursues a transatlantic sovereign AI merger, and Nvidia backs high‑end AI storage infrastructure.
Big Tech
DeepSeek debuts V4 preview with 1M context and lower costs
DeepSeek is a Chinese AI lab; it rolls out two preview versions of its newest large language model—V4 Flash and V4 Pro—designed to handle very long inputs and run more efficiently. Both are mixture‑of‑experts models with 1 million‑token context windows; V4 Pro totals 1.6 trillion parameters with 49 billion active per request, and V4 Flash totals 284 billion with 13 billion active. DeepSeek positions these as major upgrades over its V3.2 series. 1
DeepSeek highlights architecture changes, including what it calls a Hybrid Attention Architecture to better retain information across long conversations, and emphasizes advancements on reasoning and agentic tasks. The lab frames V4 as a top open‑weight option and a direct challenge to closed rivals. 2
On performance, DeepSeek says the V4 lineup nearly “closes the gap” with current leaders on reasoning benchmarks and that coding results are “comparable to GPT‑5.4,” while also noting a lag behind GPT‑5.4 and Google’s Gemini 3.1 Pro on knowledge tests—implying a several‑month trailing development curve. 1
Pricing is a headline change: V4 Flash lists at $0.14 per million input tokens and $0.28 per million output tokens; V4 Pro at $0.145 per million input and $3.48 per million output—undercutting many frontier‑class models. The launch lands amid political scrutiny: U.S. officials accuse China of large‑scale IP theft, and DeepSeek faces separate distillation allegations from Anthropic and OpenAI. 1
Industry & Biz
DeepSeek’s V4 arrives as it seeks outside funding
The release ends months of quiet from one of China’s most‑watched AI labs and comes as DeepSeek faces rising compute and talent costs and, per reporting, seeks its first round of external fundraising—factors that could influence how investors value the startup. 3
The drop follows a year after DeepSeek rattled U.S. rivals with its R1 reasoning model, and it continues China’s push to publish strong open‑weight systems. The company signals it intends to open source V4, a strategy that helped Chinese models gain global usage share through 2025. 4
For teams choosing vendors, the combination of lower usage prices and open weights positions DeepSeek as a procurement option alongside established U.S. providers—especially for long‑document and code‑heavy workflows. WSJ notes the timing could shape investor perception as the firm engages potential backers. 3
Cohere to merge with Aleph Alpha, targeting sovereign AI
Cohere, a Canada‑based enterprise AI provider, says it will merge with Germany’s Aleph Alpha to build a “transatlantic AI powerhouse.” Reports peg the combined company’s valuation around $20 billion, with Germany’s Schwarz Group committing $600 million to Cohere’s Series E. The stated aim: offer businesses and governments an alternative with more control over data and deployment. 5
Reuters reporting adds that Cohere shareholders are set to own about 90% of the combined entity and Aleph Alpha’s around 10%, and quotes Cohere’s CEO describing a push toward more secure, sovereign technology. The firms emphasize regulated sectors spanning energy, defense, finance, telecom, healthcare, and the public sector. 6
Analysts frame the deal as Europe’s sovereign AI ambitions “changing hands,” shifting a German standard‑bearer into a larger North American–European platform. The message to buyers: expect more regional deployment options, data‑residency assurances, and customized models for strict compliance environments. 7
Nvidia joins Vast Data’s $1B round at $30B valuation
Vast Data raises about $1 billion in a Series F round valuing the company at $30 billion—more than triple its late‑2023 mark. Drive Capital and Access Industries lead, with participation from Fidelity, NEA, and Nvidia, underscoring investor conviction in storage and data platforms built for AI workloads. 8
Vast reports over $4 billion in cumulative bookings, more than $500 million in committed annual recurring revenue, and positive margins and free cash flow. The company pitches an “AI operating system” approach that converges data storage, compute, and real‑time processing. 9
Why it matters: AI teams increasingly hit data bottlenecks before GPU limits; Nvidia’s participation signals a broader stack strategy reaching into data platforms. In parallel, Nvidia’s developer blog showcases tooling like FLARE for easier federated learning—another move aimed at real‑world, compliant AI deployment. 10
New Tools
DeepSeek says V4 narrows the gap on reasoning
This is a pair of large language model previews—V4 Flash and V4 Pro—built to handle very long inputs and lower usage costs. Both use a mixture‑of‑experts design to activate only a subset of parameters per request and support 1 million‑token context windows, which can fit entire codebases or long PDFs into a single prompt. 1
Pricing is unusually low for models in this performance class: V4 Flash is $0.14 per million input tokens and $0.28 per million output tokens; V4 Pro is $0.145 per million input and $3.48 per million output. DeepSeek emphasizes coding and reasoning performance but notes V4 is text‑only, unlike many closed peers that handle images, audio, or video. 1
If your work leans on long‑form analysis, summarizing legal or policy documents, or large‑repo code assistance, the 1M context and pricing may shift your model mix. For encyclopedic Q&A across niche facts, DeepSeek acknowledges knowledge trailing leading closed models on some tests. 1
What This Means for You
Lower token prices plus 1M‑token context make it feasible to run end‑to‑end “document in, decision out” workflows—think 200‑page decks, full contracts, or multi‑service codebases—without elaborate chunking. That can simplify ops and reduce glue‑code maintenance for teams that currently stitch retrieval pipelines around shorter‑context models. 1
For buyers in regulated markets, the Cohere–Aleph Alpha tie‑up points to more “sovereign AI” options where data locality, on‑premise deployments, and customized safeguards are contractable requirements rather than afterthoughts. That can shorten procurement cycles for ministries, banks, and critical infrastructure. 5
Back‑end readiness matters as much as models: Vast Data’s raise at a $30 billion valuation highlights that storage bandwidth and data orchestration frequently gate AI throughput. If GPUs idle while storage feeds them, your unit economics suffer—an argument to involve infra leads early in AI rollouts. 9
Model selection should match content: V4 is text‑only and optimized for long‑context reasoning and code. If your tasks need images, audio, or video, plan a hybrid stack. If you primarily analyze sprawling text, the economics of V4 Flash/Pro may be compelling. 1
Action Items
- Run a 1M‑context pilot: Take one long document or codebase your team struggles with and test it end‑to‑end on DeepSeek V4 Flash/Pro. Track quality and total token cost.
- Estimate monthly savings: Using V4’s listed prices, model your current prompt/response volumes to compare projected spend versus your incumbent model.
- Request a sovereign‑AI brief: If you work in a regulated sector, ask your Cohere or Aleph Alpha rep for deployment options around data residency and on‑prem controls.
- Check storage throughput: Sit down with infra/data leads to map where storage or networking may bottleneck GPU jobs before expanding AI workloads.
Comments (0)