DeepSeek lowered the price floor for open-weight AI APIs. The company cut V4-Pro prices by 75% and reduced cache-hit fees by 10x across its lineup. This move adds pressure on US API pricing. In parallel, Microsoft completed Copilot rollout to all 743,000 Accenture employees. Microsoft says this is its largest enterprise Copilot deployment to date.
Capital is still concentrating around senior researchers leaving frontier labs. Ineffable Intelligence, founded by DeepMind's former reinforcement learning lead, raised $1.1 billion at a $5.1 billion valuation. The raise extends the pattern of pre-product labs attracting frontier-scale institutional capital.
DeepSeek cut V4-Pro API prices 75% and reduced cache-hit fees by 10x.
DeepSeek announced a 75% promotional discount on V4-Pro through May 5. The company also permanently cut cache-hit fees to one-tenth of prior rates across its lineup.
This targets agent workloads that repeat similar queries. In those workloads, cache-hit pricing often drives total cost more than the headline token rate.
DeepSeek's standard rates already undercut OpenAI, Google, and Anthropic for these models. The cache-hit cut increases that cost gap for agent-heavy usage.
Microsoft completed Copilot rollout to all 743,000 Accenture employees.
Microsoft completed Microsoft 365 Copilot rollout to about 743,000 Accenture employees in 120 countries. The deployment started with a pilot of a few hundred senior leaders in August 2023.
This shows a large enterprise moving from pilot to full workforce deployment. The timeline from pilot to full rollout was under three years.
Microsoft reported metrics from a survey of about 200,000 users: 97% said they completed routine tasks up to 15 times faster, and 89% were monthly active users. These are Microsoft-published figures from a self-selected cohort.
Former DeepMind RL lead raised $1.1 billion for Ineffable at a $5.1 billion valuation.
David Silver, former lead of Google DeepMind's reinforcement learning team, raised a $1.1 billion seed round for Ineffable Intelligence. The UK-based lab was founded in late 2025. Sequoia and Lightspeed co-led the round, with participation from Nvidia, Google, and the UK's Sovereign AI Fund.
The raise fits a broader pattern: senior researchers leave major labs and attract large early funding. Capital is concentrating around people and training approaches, not only around incumbent frontier labs.
Ineffable focuses on reinforcement learning from experience rather than human-generated data. Comparable activity is also visible: Recursive Superintelligence is reportedly raising up to $1 billion, and AMI Labs raised $1 billion in March 2026.
DeepSeek V4-Pro
worth attentionDeepSeek permanently cut cache-hit fees to one-tenth of prior rates. This matters for agent workloads, where repeated queries can make cache pricing the main cost driver.
Use this for agent workloads that repeat similar prompts on V4-Pro or V4-Flash. At high query volume, the cache-hit fee change can matter more than the 75% promotional discount.
Source: Yahoo Finance (Reuters) →No friction point identified this issue.
$8 billion
Cadence Design Systems reported a record EDA backlog. The company expects half to convert to revenue in the next twelve months. This is contracted demand for AI accelerator design work, not a projection.
Source: MarketScreener (Reuters) →Two data points show that AI infrastructure demand has moved from projection to contracted reality. One is an $8 billion EDA backlog with half expected to convert in the next year. The other is a completed 743,000-seat enterprise Copilot deployment. Against that baseline, DeepSeek's permanent cache-hit cut targets the cost pressure that agent systems feel at scale.
At the same time, investors are funding pre-product reinforcement learning labs at multi-billion-dollar valuations. That pattern signals a bet that current training paradigms may have limits, even as infrastructure commitments continue to rise.