The Adept talent transfer and investor recoupment
Amazon hired the majority of Adept’s researchers in June. This leaves Adept with approximately one third of its staff. Adept reached a \$1 billion valuation after raising \$414 million from investors including Greylock, General Catalyst, Nvidia, Spark Capital, Atlassian, Microsoft, Workday, Frontiers Capital, PSP Partners, and SV Angel. Earlier rounds included a \$62.88 million Series A and a \$1 million Series A-1 in April 2022, which included contributions from Andrej Karpathy, Chris Re, Howie Liu, Root Ventures, Sarah Meyohas, and Scott Belskey. While Amazon did not buy the company, the arrangement allows investors like Greylock and General Catalyst to recoup their \$414 million investment through a \$25 million payment overseen by co-founder David Luan. I find the \$25 million payout for a \$414 million investment to be an incredibly lean recovery for the original venture capitalists. The US Federal Trade Commission sent a list of questions to Amazon to investigate whether the mass hiring of Adept researchers constitutes an acquisition that bypasses the merger notification rules used to regulate large technology deals. This scrutiny follows a similar investigation into Microsoft’s March hiring of Inflection AI staff, where Microsoft paid \$650 million in licensing fees. Adept builds technology that interprets natural language instructions to perform actions across websites and business applications. Its ACT 1 Action Transformer processes instructions by using visual perception to identify buttons, fields, and menus on a screen to execute workflows like typing or scrolling.
| Adept Funding Round | Date | Amount | Post-Money Valuation |
|---|---|---|---|
| Series B | 03/15/2023 | \$200M | \$1B |
| Series B-1 | 03/15/2023 | \$2M | \$1B |
| Series A | 04/26/2022 | \$62.88M | \$182.68M |
| Series A-1 | 04/26/2022 | \$1M | \$182.68M |
Amazon scales infrastructure while cutting AGI roles
Amazon forecasts \$200 billion in annual capital expenditure for 2026 to support its AI expansion. This massive budget allocation occurs while the company reduces staff within its artificial general intelligence organization. Amazon eliminated roles in model customization, post-training, data services, and information. These cuts follow the departure of Rohit Prasad and the consolidation of AGI work under Senior Vice President Peter DeSantis in December 2025. David Luan also left the AGI Lab in February after joining through Amazon’s hiring of Adept employees. I see a clear conflict between Amazon’s push for massive computing infrastructure and its decision to eliminate the very roles that customize the models. Senior Vice President Peter DeSantis acknowledged that Amazon’s models haven’t been at the very frontier for the very largest, most demanding workloads. Workers reporting to AGI data services executive Adeeb Shanaa and AGI information executive Vishal Sharma also reported being impacted by the restructuring. Amazon is raising tens of billions of dollars in debt to support its infrastructure build-out, which represents a 50% increase in capital expenditure compared to 2025. The company released its Nova family of foundation models in 2024. Amazon has eliminated roughly 30,000 roles since October, including 16,000 jobs in January across AWS, retail, Prime Video, and human resources. Affected US employees receive 90 days of pay and benefits alongside transitional healthcare support.
The move toward agentic automation
Adept moved away from general intelligence breakthroughs to focus on products for businesses because the cost of training models became too high. The company now focuses on agentic AI, which uses autonomous, goal-seeking behavior to manage complex business processes. This differs from generative AI, which focuses on content creation, and traditional AI, which follows rule-based systems. Agentic AI maintains objectives over time and adjusts strategies based on environmental feedback. You likely know that enterprise leaders prefer tools that solve specific, high-value problems rather than general tools. Salesforce demonstrated this approach by acquiring Qualified to improve decision velocity at the moment of buyer intent. In the enterprise space, usage often remains metered even with premium subscriptions. For example, UiPath allows agentic AI usage within an attended enterprise license, but users must manage AI units or PlatformUnits. Each user receives a daily free quota of LLM calls for design or testing, but production usage consumes purchased units. High-volume tasks require sufficient PlatformUnits to avoid exhausting the subscription. Companies like Webflow suggest that the next era of growth depends on discoverability and conversion rather than traditional sales funnels. To succeed, revenue teams must track pipeline conversion, average deal size, sales cycle length, and win rate by segment. They must also utilize a rolling ICP that evolves based on product usage patterns and customer conversations. How will Amazon maintain its frontier model status while cutting the very researchers who build them?
