Amazon's AI Overhaul Signals a Company at War with Itself

On December 17, 2025, Andy Jassy sent an internal memo announcing that Amazon had reached an “inflection point” requiring a renewed focus on several key technologies. The language was optimistic, forward-looking, corporate-speak of the highest order. What actually happened was something more revealing: Amazon’s artificial intelligence chief Rohit Prasad is leaving at the end of the year, the company is consolidating its scattered AI efforts under a single leader, and the architect of Amazon’s cloud infrastructure is now running the show. The framing says opportunity. The substance says crisis management.

The restructuring places Peter DeSantis, a 27-year Amazon veteran who launched EC2 and oversaw the Annapurna Labs chip acquisition, at the helm of a unified group encompassing AI model development, custom silicon, and quantum computing. CNBC reports that DeSantis will report directly to Jassy—a signal of how seriously leadership is treating the AI gap. Meanwhile, Pieter Abbeel, the UC Berkeley professor and former OpenAI researcher who joined Amazon through its Covariant acquisition, will lead the frontier model research team. The moves happened two weeks after Amazon unveiled its Nova 2 models at re:Invent , models that earned praise for efficiency but still trail OpenAI’s GPT offerings, Anthropic’s Claude, and Google’s Gemini on the metrics that matter.

The timing is not coincidental. Amazon is simultaneously in talks to invest over $10 billion in OpenAI, a deal that would require the ChatGPT maker to use Amazon’s Trainium chips and rent additional data center capacity from AWS. This is the same Amazon that has already poured $8 billion into Anthropic, making it the primary cloud provider for one of OpenAI’s fiercest competitors. The company is hedging so aggressively that it risks funding both sides of the AI race while its own models languish in the middle of the pack. Something had to change, and the reorganization is Jassy’s answer—an admission, wrapped in optimistic language, that Amazon’s original AI strategy did not work.

The question worth asking is whether structural changes can solve what may be a cultural problem. Amazon built the cloud computing industry through relentless customer obsession and operational excellence. Those virtues produced AWS, the most profitable division of the company. But they have not produced a competitive large language model, a voice assistant that can match ChatGPT, or a chip ecosystem that developers choose voluntarily over NVIDIA. The inflection point Jassy identifies is real, but it may require more than a reorg to navigate.

The Prasad problem and why Alexa never got its upgrade

Rohit Prasad joined Amazon in 2013, during the early days of Alexa. He rose through the ranks as the voice assistant grew from novelty to household name, eventually becoming senior vice president and head scientist for artificial general intelligence in mid-2023. His portfolio was ambitious: lead the team building Amazon’s foundation models and position the company to compete with OpenAI and Google at the frontier. It did not work out that way.

The evidence of failure is hard to miss. Fortune reported that Amazon was “caught flat-footed” by ChatGPT’s November 2022 launch, leading to a “frantic, frenetic few months” as the Alexa organization struggled to coalesce around a vision for generative AI. The project to upgrade Alexa with large language model capabilities, codenamed “Remarkable Alexa” or “Project Banyan,” was supposed to launch in October 2024. It slipped repeatedly. Beta testers reported irrelevant responses, hallucinations, and unreliable smart home integration. WinBuzzer noted that the technical challenges stemmed from embedding large language models into Alexa’s existing architecture—a problem that should have been anticipated years earlier.

When the new Alexa finally launched in limited form in early 2025, Amazon decided to charge consumers $20 per month for access, though Prime members would get it free. TheStreet’s verdict was blunt: “While competitors are finding ways to weave AI into nearly everything they do, Amazon is still struggling to make Alexa sound competent.” The company has a multibillion-dollar stake in Anthropic, the company behind Claude. It has hundreds of millions of Echo devices in homes worldwide. Yet Alexa+ still feels like it belongs to a different era.

The numbers tell the same story. According to eMarketer, Alexa has roughly 77.6 million U.S. users in 2025—trailing Google Assistant at 92.4 million and Apple’s Siri at 87 million. Amazon was first to market with a smart speaker, first to achieve mass adoption, and first to have a voice AI presence in kitchens and bedrooms across America. That lead evaporated. Part of the problem, Fortune reported, was internal politics: “Privacy concerns have kept Alexa’s teams from using Anthropic’s Claude model, former employees say—but so too have Amazon’s ego-driven internal politics.” A company that prides itself on customer obsession let organizational dysfunction get in the way of shipping a competitive product.

Prasad’s departure is framed as voluntary—he’s “leaving at the end of the year”—but the context makes clear that his tenure did not produce what Amazon needed. The Nova models his team built are efficient and cost-competitive, but efficiency is table stakes when OpenAI and Anthropic are pushing the frontier of what AI can do. Amazon needed breakthrough capability. It got incremental improvement. That gap explains why DeSantis, an infrastructure expert, is now running AI development, and why Abbeel, a robotics and reinforcement learning specialist, is leading frontier research. The implicit message is that the previous approach—Prasad’s approach—did not deliver.

The Apple playbook and why DeSantis might actually make sense

The appointment of Peter DeSantis is the most significant signal in the reorganization. DeSantis is not an AI researcher. He’s an infrastructure guy—the person who built EC2, the compute backbone of AWS, and who oversaw the acquisition of Annapurna Labs, the Israeli chip design company that created Amazon’s Graviton processors and the Trainium AI accelerators. His expertise is in making things work at scale, integrating hardware and software, and squeezing efficiency out of systems. That skillset may be exactly what Amazon needs.

Jassy’s memo, published by Amazon, specifically calls out “the advantages of optimizing across models, chips, and cloud software and infrastructure.” This is Apple’s playbook for AI—control the silicon and the models together, optimize end-to-end, create a vertically integrated stack that competitors using commodity hardware cannot match. The Register observed that this is “the strategic bet Amazon is making here.” If it works, Amazon could offer AI services at price points that OpenAI and Anthropic, dependent on NVIDIA GPUs, cannot match.

The Trainium chip roadmap supports this thesis. At re:Invent 2025, Amazon unveiled Trainium3, a 3-nanometer chip that Amazon claims delivers 4.4x more compute performance, 4x greater energy efficiency, and nearly 4x more memory bandwidth than its predecessor. Customers including Anthropic, Karakuri, Metagenomi, and Splash Music are reportedly reducing training and inference costs by up to 50% with Trainium, while Decart claims 4x faster inference for real-time generative video at half the cost of GPUs. The company also previewed Trainium4, which will support NVIDIA’s NVLink Fusion interconnect—a hedge that allows Trainium systems to work alongside NVIDIA hardware rather than requiring customers to choose one or the other.

The challenge is adoption. SiliconANGLE noted that Amazon faces a key obstacle: “its silicon still lacks the deep and mature software ecosystem that makes Nvidia’s GPUs so easy to adopt.” NVIDIA’s CUDA platform has a decade of developer tools, libraries, and documentation behind it. Trainium has compelling benchmarks and Anthropic as an anchor customer, but winning over the broader developer community requires more than raw performance numbers. Developers choose ecosystems, not chips, and Amazon’s ecosystem is newer and less proven.

DeSantis’s track record suggests he understands this. EC2 succeeded not just because Amazon had cheap compute, but because Amazon made that compute easy to use—APIs, documentation, integrations with everything else in AWS. GeekWire reported that DeSantis “launched Amazon EC2, the company’s core cloud computing infrastructure, oversaw the acquisition of chip designer Annapurna Labs in 2015, and most recently ran AWS Utility Computing.” That’s a career building platforms, not products. If Amazon’s AI problem is one of integration—making Trainium, Nova, and Bedrock work together seamlessly—then DeSantis is a logical choice. If the problem is one of research breakthrough—catching up to GPT-5.2 and Claude Opus 4.5 on reasoning, coding, and general intelligence—then an infrastructure executive may be solving the wrong problem.

The Pieter Abbeel appointment addresses the research gap, at least in theory. Abbeel won the 2021 ACM Prize in Computing for foundational work in robot learning, co-founded Covariant to apply that research to warehouse automation, and joined Amazon when Covariant was acquired in 2024. His background includes a stint at OpenAI from 2016-2017, giving him direct exposure to frontier model development. Amazon Science lists his research interests as generative AI, reinforcement learning, and humanoid robotics. If anyone can help Amazon catch up on the model side, Abbeel has the credentials. The question is whether Amazon’s organizational culture will let him move fast enough.

The OpenAI hedge and what $10 billion really buys

The timing of the leadership reorganization coincides with reports that Amazon is in talks to invest at least $10 billion in OpenAI. CNBC reported that the deal would value OpenAI above $500 billion and require the AI lab to use Amazon’s Trainium chips and rent additional AWS data center capacity. This comes on top of the $38 billion that OpenAI has already committed to renting servers from AWS over the next seven years. If the deal closes, Amazon would have major financial stakes in both OpenAI and Anthropic—the two leading proprietary AI labs competing for the same customers.

The logic, from Amazon’s perspective, is straightforward: if you cannot beat them, invest in them, and make sure they build on your infrastructure regardless of who wins. Engadget noted that the deal would also include OpenAI helping Amazon with its online marketplace, similar to partnerships the AI lab has with Etsy, Shopify, and Instacart. Microsoft holds exclusive rights to OpenAI’s most advanced models until the 2030s, so Amazon would not be able to offer those models through Bedrock. But access to Trainium revenue and AWS compute spending could be worth the investment even without model distribution rights.

The hedge strategy is not inherently flawed, but it does highlight a tension in Amazon’s approach. TechCrunch observed that these “circular deals” have become popular in the AI industry—cloud providers investing in AI labs that then spend those investments on cloud services. The money flows in a loop, creating the appearance of massive investment while the actual capital efficiency is debatable. Amazon gave Anthropic $8 billion; Anthropic spends a significant portion on AWS. Amazon may give OpenAI $10 billion; OpenAI will spend a significant portion on AWS. The investments are real, but the net capital deployed to AI development is less than the headline numbers suggest.

More critically, the willingness to fund competitors signals a lack of confidence in Amazon’s own AI capabilities. If Nova models were competitive with GPT-5 and Claude Opus, Amazon would not need to invest billions in the companies that build superior alternatives. The investments are insurance policies—acknowledgments that Amazon’s in-house efforts may not be enough to keep AWS relevant as AI becomes the primary driver of cloud demand. 24/7 Wall St. argued that the OpenAI investment could be a waste because Microsoft’s exclusive access to the best models means Amazon gets relatively little for its money. That may be too pessimistic, but the criticism points to a real problem: Amazon is spending billions to remain dependent on external AI providers rather than building capability internally.

The counterargument is that Amazon does not need to win the model race to win the AI business. AWS already offers Bedrock, a managed service that provides access to foundation models from Anthropic, Meta, Mistral, Cohere, and Amazon itself. If enterprise customers want to use Claude or Llama, they can do so through AWS infrastructure. CNBC reported that Amazon released Nova Act, an AI agent model designed for browser automation, positioning it against similar offerings from OpenAI and Anthropic. The platform strategy does not require Amazon to have the best model—just a good-enough model combined with the best infrastructure. Whether that’s sufficient depends on how quickly AI capabilities become commoditized and whether the model layer captures more value than the infrastructure layer over time.

The infrastructure investment is substantial. Amazon opened an $11 billion AI data center in Indiana in October 2025—Project Rainier, named for the mountain and dedicated to training and running models from Anthropic. The company claims the finished facility will be “the world’s largest AI compute cluster,” larger even than the Stargate project that OpenAI and Microsoft have proposed. That physical footprint represents Amazon’s long-term bet: even if OpenAI builds the best models, those models need somewhere to run, and AWS wants to be that somewhere. The combination of proprietary chips, massive data centers, and investments in multiple AI labs creates a diversified portfolio. The risk is that diversification becomes fragmentation, and Amazon ends up funding everyone else’s AI breakthroughs while its own Nova models remain also-rans.

The ways this bet could blow up—and what matters now

The bullish case for Amazon’s reorganization is that vertical integration will create a cost advantage that compounds over time. If Trainium chips improve faster than NVIDIA GPUs, if DeSantis can integrate models and silicon the way Apple integrates its chips and software, and if AWS’s scale makes Amazon the low-cost provider of AI compute, then the current model gap becomes irrelevant. Enterprise customers will choose AWS because it’s cheaper, more convenient, and tightly integrated with the cloud services they already use. The billions invested in OpenAI and Anthropic ensure that AWS remains the infrastructure of choice regardless of which lab produces the best model.

The bearish case is that model capability matters more than infrastructure efficiency, and Amazon is restructuring deck chairs while the ship takes on water. OpenAI, Anthropic, and Google are pushing the frontier of what AI can do—reasoning, coding, scientific research, autonomous agents. Amazon’s Nova models, for all their efficiency, do not compete at that frontier. If the most valuable AI applications require frontier-level intelligence, then cost-competitive inference on second-tier models is not a winning proposition. Customers will pay the premium for GPT-5.2 or Claude Opus because those models can do things Nova cannot.

The historical precedent cuts both ways. Fortune noted that Amazon faced skepticism when AWS launched in 2006, and the company proved doubters wrong by building the dominant cloud platform through relentless execution. But Amazon has also had notable failures—the Fire Phone, Amazon Restaurants, Amazon Destinations—where execution alone was not enough to overcome product-market fit problems. The AI reorganization is a bet that execution can close the capability gap. Whether that bet pays off depends on factors that organizational charts cannot control: the pace of model research, the trajectory of chip development, and the preferences of developers who increasingly have multiple cloud providers competing for their business.

The market signal worth watching is Anthropic’s behavior. Amazon announced an additional $4 billion investment in Anthropic in 2024, bringing total investment to $8 billion. Anthropic is training its models on Trainium and running inference on AWS. That partnership is the best evidence that Amazon’s silicon can compete with NVIDIA at the highest levels of AI development. If Anthropic’s next-generation models continue to match or exceed OpenAI’s, and if those models are trained on Trainium, then Amazon’s vertical integration thesis gains credibility. If Anthropic quietly shifts more workloads to NVIDIA, that tells a different story.

For developers and enterprises evaluating cloud AI providers, the reorganization creates uncertainty but also opportunity. AWS is clearly prioritizing AI investment, and the Trainium roadmap promises substantial cost reductions for inference and training workloads. The platform bet—Bedrock offering multiple models through a unified API—remains sound even if Nova models underperform. The question is whether to bet on Amazon’s execution capability or to stay with NVIDIA-based infrastructure where the software ecosystem is more mature and the model options more proven. There is no obviously correct answer; the choice depends on workload characteristics, cost sensitivity, and tolerance for platform risk.

What the reorganization does not address is the cultural question. Amazon’s AI struggles were not primarily technical—they were organizational. Internal politics prevented Alexa from using Anthropic’s Claude. Competing priorities fragmented AI development across multiple teams. The original AGI organization, created in 2023, did not produce breakthrough results despite significant investment. DeSantis can consolidate reporting lines and integrate silicon with software, but changing how Amazon makes decisions about AI products requires more than a memo. The inflection point Jassy identifies is real, but inflection points are moments of maximum instability. Amazon has made its bet. The next twelve months will show whether it was the right one.

The subtext of every corporate reorganization is an admission that the previous structure did not work. Amazon wrapped that admission in optimistic language about inflection points and long-term potential, but the message is clear: Prasad’s approach failed, the model gap persists, and the company is trying something different. Whether DeSantis and Abbeel can succeed where Prasad did not depends on execution, timing, and a fair amount of luck. Amazon has the resources to win the AI race. The question is whether it has the culture—and after this week, whether it has finally organized itself to use those resources effectively.

For operators watching Amazon’s AI strategy, a few principles emerge. First, vertical integration matters: controlling chips and models together creates optimization opportunities that horizontal competitors cannot match, and Amazon is explicitly pursuing that Apple-style playbook. Second, platform bets can succeed even without the best individual product: AWS does not need Nova to beat GPT-5 if Bedrock offers convenient access to Claude and Llama alongside cost-effective inference on Amazon silicon. Third, organizational structure reflects strategic priorities, and Amazon’s decision to put an infrastructure veteran in charge of AI signals that efficiency and integration matter more right now than research breakthroughs. Fourth, hedging has limits: investing in OpenAI and Anthropic provides insurance but also signals a lack of confidence in internal capabilities, and at some point Amazon must deliver its own competitive models or accept permanent dependency on external providers.

The AI race is far from over, and Amazon’s resources ensure it will remain a major player regardless of any single quarter’s results. But the reorganization is a reset, not a victory lap. Jassy’s inflection point memo acknowledges that Amazon needs to change course. The next chapter will show whether that change is enough.