SerpApi.Org
Amazon confirmed Wednesday that it eliminated an undisclosed number of positions inside its artificial general intelligence organization — the division responsible for the Nova foundation models, custom silicon, and quantum computing programs — in a move that crystallizes a deliberate strategic wager: the company believes that deploying AI to enterprise customers will deliver faster returns than researching frontier models.
The roles reportedly eliminated were concentrated in model customization and post-training, according to posts on professional networking sites cited by CNBC — precisely the work involved in shaping how a language model behaves, aligns to user intent, and adapts to specialized domains. That is not a coincidence. Six weeks before the cuts, Amazon launched Nova Forge, a product specifically designed to let enterprise customers perform their own model customization and post-training — effectively automating and externalizing the function the cut roles performed manually.
Amazon simultaneously holds more than 100 open positions within the same AGI unit, signaling that the company intends to keep recruiting for roles it considers strategically essential — the research it is cutting is not the research it is hiring for.
Amazon's AGI Bet Is Now on Deployment, Not Research
In a statement to CNBC, an Amazon spokesperson said the company is "sharpening its focus on the initiatives that matter most for customers," adding that this required "some difficult decisions, including eliminating some roles within parts of our AGI organization."
Three weeks before the AGI cuts, AWS announced a $1 billion initiative to embed its engineers directly with enterprise customers building agentic AI systems — a forward-deployed model designed to accelerate real-world adoption. Read those two moves together, and the logic is consistent: Amazon is moving its AI effort closer to the buyer, not the research frontier.
AWS already generates an annualized revenue run rate above $15 billion specifically from AI services — a figure CEO Andy Jassy disclosed for the first time in his April 2026 shareholder letter. That number represents roughly 10% of AWS's total $142 billion annualized run rate, and Jassy described the trajectory as "ascending rapidly."
Who Was Cut and What They Did
According to reporting citing Reuters, employees under Adeeb Shanaa, vice president of AGI data services, and Vishal Sharma, vice president of AGI information, were among those who reported being affected by the cuts, based on posts to professional networking forums. Amazon declined to confirm which sub-teams were involved or how many positions were eliminated.
Affected U.S.-based workers were told they would receive 90 days of pay and benefits, access to outplacement services and transitional healthcare, and remain eligible for severance.
The cuts were concentrated in post-training and model customization — work that happens after a model's computationally expensive pretraining phase is complete. Post-training is where behavioral alignment occurs: instruction-following, safety guardrails, and the reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) techniques that shape how a model actually responds to users. Cutting these roles while keeping infrastructure investment suggests Amazon has decided that behavioral refinement is becoming a product capability — something customers configure themselves through tools like Nova Forge — rather than an internal research function.
Nova Forge and the Automation of Model Customization
Amazon launched Nova Forge in December 2025, describing it as giving organizations "unprecedented access to Nova model training, enabling them to build custom frontier models that deeply embed domain expertise without the traditional barriers of cost, compute, and time." The product allows enterprise customers to perform continued pre-training, supervised fine-tuning, and direct preference optimization on their own data — the three core techniques of model customization that were the specialty of the roles now reportedly being eliminated.
This is not a coincidence the draft of record glossed over. If Nova Forge succeeds at democratizing model customization, the internal humans who performed that function manually become redundant. Amazon is not abandoning model customization. It is externalizing it to customers, charging for compute on Bedrock and Trainium chips, and reducing the human capital required to support it internally. The layoff and the product launch are the same strategy viewed from two angles.
The $200 Billion Bet and Why It Requires Cutting, Not Just Spending
Amazon's capex guidance for 2026 stands at approximately $200 billion — a figure Jassy stated on the Q4 2025 earnings call in February, directed predominantly at AWS data centers, networking, and custom silicon. This represents a roughly 52% increase over the full-year 2025 capex of $131.8 billion. To help finance that buildout, Amazon has raised tens of billions of dollars in bond markets.
AWS posted its fastest growth rate in 15 quarters in Q1 2026, with segment sales rising 28% year over year to $37.6 billion at a 37.7% operating margin. Q1 capital expenditure alone reached $43.2 billion, compressing near-term free cash flow.
The engine behind the infrastructure bet is Amazon's custom silicon business. Trainium3 — built on TSMC's 3-nanometer process node — delivers 2.52 petaflops of FP8 compute per chip and 144 gigabytes of HBM3e memory. A full 144-chip UltraServer delivers 362 MXFP8 petaflops of aggregate compute — performance that, according to third-party analysis, matches NVIDIA's Blackwell NVL72 at approximately 50% lower total cost of ownership. Jassy confirmed on the Q1 2026 earnings call that Trainium3 started shipping in 2026, is 30–40% more price-performant than Trainium2, and is already nearly fully subscribed.
The strategic logic is explicit in DeSantis's own words. Speaking at VivaTech Paris in June 2026, he described chip and model development as requiring lockstep co-design: DeSantis said that "if the chips are not telling the model designers what capabilities are coming and where they can optimize, then we're not doing the science necessary to take advantage of those capabilities until the chips are available, and then you're waiting months and months." That means the infrastructure executive running the AI division is not a mismatch — it is the point. Amazon's AI strategy is an infrastructure strategy, and the human capital that does not fit that frame is what is being shed.
Amazon's Acknowledged Gap at the Frontier
The AGI cuts arrive against a backdrop of significant leadership churn and a candid admission about competitive position. At a CNBC interview in June 2026 — one month before the cuts — DeSantis stated plainly that "it's a fair narrative that our models haven't been at the very frontier for the very largest, most demanding workloads." He expressed hope that Amazon would be "in the conversation about leading models in the coming year."
The leadership exits have been striking. Rohit Prasad, who had led Amazon's AGI efforts, departed at the end of 2025. In December 2025, Amazon consolidated its AGI work under Peter DeSantis, a 27-year Amazon veteran who previously led AWS infrastructure. In February 2026, David Luan — who had joined Amazon through the acquihire of AI startup Adept in 2024 and had led the AGI Lab in San Francisco — also left the company. The AGI lab subsequently shifted to report directly to DeSantis.
Independent benchmark comparisons still favor rival offerings over Nova 2, Amazon's latest flagship model, which launched in December 2025 and has approximately 50,000 customers. Amazon's own technical report acknowledges that Nova 2 delivers "strong performance across a broad range of enterprise workloads" while positioning it on price-performance rather than raw capability against GPT-class models.
Amazon also holds a significant strategic investment in Anthropic — the AI lab that competes directly with Nova on the same Bedrock platform. Amazon's Q1 2026 net income included a $16.8 billion pre-tax gain from the Anthropic investment. The arrangement creates an unusual competitive structure: Amazon profits from Anthropic's success even as DeSantis rebuilds the internal model team to challenge it.
Part of Amazon's Largest Corporate Workforce Contraction on Record
The AGI cuts are the latest chapter in an extended corporate workforce reduction at Amazon. The company eliminated more than 30,000 corporate roles since October 2025 — approximately 10% of its corporate staff and the largest workforce reduction in its three-decade history, surpassing the roughly 27,000 jobs cut between late 2022 and early 2023.
Amazon Senior Vice President of People Experience Beth Galetti, announcing the January round, said the company was working to reduce management layers, increase ownership, and strip out bureaucracy, pledging continued hiring in what she described as "strategic areas and functions that are critical to our future."
CEO Andy Jassy has been direct about the long-term trajectory. Speaking at the World Economic Forum in Davos around the time of the January reductions, Jassy said AI would affect employment over time and that he could envision the company operating with a leaner corporate headcount in the years ahead.
Is Amazon Winning or Retreating?
The apparent contradiction of cutting AI researchers while setting a $200 billion infrastructure record resolves when viewed through the lens of what Amazon is actually selling. AWS is not primarily selling model intelligence — it is selling compute capacity, managed inference, and the platform on which customers bring their own models or choose from a marketplace. Amazon's Bedrock platform distributes Nova models alongside third-party options including Anthropic's Claude, Meta's Llama, and DeepSeek.
In that architecture, Amazon profits from inference regardless of which model wins. The Trainium chip — not the Nova model — is the irreplaceable asset. DeSantis himself has described the business in exactly these terms: cutting-edge silicon that forces model design to improve in lockstep, rather than model research that runs ahead of the chip. Cutting post-training researchers while keeping chip engineers is internally consistent within that thesis.
Amazon is scheduled to report second-quarter 2026 results on July 30, with management guidance calling for net sales of $194 billion to $199 billion and operating income of $20 billion to $24 billion. Bank of America raised its AWS revenue growth forecast for the quarter to 33% year over year — and estimated that Anthropic-related workloads alone could contribute more than $1.5 billion in sequential AWS revenue growth — underscoring that Amazon's AI revenue engine runs whether or not its models lead at the frontier.
The question the AGI layoffs raise is not whether Amazon is retreating from AI — it clearly is not, as the $200 billion spending plan and open recruitment signal. The question is whether a company that has publicly acknowledged its models are not at the frontier, and that is now also reducing the human researchers tasked with improving them, can close that gap through infrastructure and customer deployment alone.
Frequently Asked Questions
Why is Amazon cutting AI researchers if it's spending $200 billion on AI?
Amazon's spending is directed primarily at infrastructure — data centers, networking, and custom Trainium chips — not at the human research that builds and refines language models. The company appears to be making a strategic bet that deploying AI to enterprise customers through tools like Bedrock and its $1 billion forward-deployed engineering program will generate faster near-term returns than pure model research. The post-training and model customization roles reportedly cut are also the exact functions that Amazon's Nova Forge product is designed to automate and externalize to customers — suggesting the layoffs and the product launch reflect the same underlying strategy.
How does Amazon's Nova compare to OpenAI and Anthropic models?
Amazon's own AI chief, Peter DeSantis, acknowledged in a June 2026 CNBC interview that it is "a fair narrative that our models haven't been at the very frontier for the very largest, most demanding workloads." Nova 2, launched in December 2025, has roughly 50,000 customers and positions itself on price-performance rather than top-tier capability against GPT-class and Claude-class models. Independent benchmark leaderboards still favor rival offerings. Amazon is also Anthropic's largest investor and distributes Anthropic's Claude on its own Bedrock platform, creating an unusual arrangement where Amazon profits from the competitor that outperforms its own models.
What is Amazon Trainium and why does it matter more than Nova right now?
Trainium is Amazon's custom AI chip, designed and built by its Annapurna Labs division. Trainium3, the current generation, is built on TSMC's 3-nanometer process and packs 2.52 petaflops of FP8 compute per chip. A 144-chip UltraServer matches NVIDIA's Blackwell NVL72 at approximately 50% lower total cost of ownership according to third-party analysis. Amazon's entire $200 billion infrastructure bet depends on Trainium undercutting NVIDIA on cost while matching it on performance — making the chip business the strategic center of gravity, not the model research. DeSantis describes chip and model design as requiring lockstep co-development, which is why an infrastructure executive now runs both.
What does this mean for enterprise teams choosing an AI cloud platform?
The strategic pivot means Amazon is increasingly a platform company that enables AI rather than a frontier AI lab that competes on model quality. That distinction matters for enterprise buyers. If your workloads require the most capable reasoning models, Amazon's Bedrock platform gives you access to Anthropic's Claude, Meta's Llama, and other third-party models — not just Nova. If your priority is cost and infrastructure scale, Amazon's Trainium chips and Bedrock managed inference may offer a cost advantage. But if you need model customization at depth, the reduction in Amazon's internal post-training team means more of that work may need to be done by your own team using Nova Forge, rather than by Amazon specialists.
Related Articles
