Character.AI Had 20 Million Users and Filed for Bankruptcy Anyway. Here's the Consumer AI Retention Math That Nobody Wanted to Face.

The Wall Street Journal reported on October 11, 2026 that Amazon is in advanced discussions to acquire Decart, a three-year-old AI startup, at a valuation of approximately $7 billion. The deal has not closed. Decart has not publicly disclosed its revenue. Amazon has not commented. And Anthropic — which had its own $6 billion acquisition discussions with Decart earlier this year before walking away after due diligence — had the same information Amazon has now and decided the price was too high.

The fact that Amazon is reportedly willing to pay more for an asset that Anthropic passed on at a lower price tells you exactly what is actually being purchased. This is not an AI capability acquisition. This is a strategic infrastructure play in the war to reduce Nvidia dependence, and the $7 billion price reflects the strategic value of that outcome to a company that spends tens of billions per year on AI compute.

What Decart Actually Does

Decart was founded in 2023 by a team of engineers with backgrounds in distributed systems and hardware optimization. Its core product, the Decart Optimization Stack (DOS), is software that allows AI model workloads to run across chips from different vendors — Nvidia GPUs, Google TPUs, and Amazon's Trainium chips — with minimal code changes required.

For context on why this is technically difficult: AI models are trained and deployed using CUDA, Nvidia's proprietary programming framework. CUDA code is optimized for Nvidia's specific hardware architecture and memory bandwidth characteristics. Moving a CUDA-trained model to Google's TPUs or Amazon's Trainium requires either significant rewriting of the model code or a translation layer that understands both architectures deeply enough to replicate the numerical precision Nvidia hardware provides. Building that translation layer is the core technical problem Decart has spent three years solving.

Beyond the DOS portability software, Decart has built two application products: Lucy, a real-time video editing and generation model, and Oasis, a world model that generates simulated environments for robotics and autonomous driving applications. Amazon has stated publicly that Decart was running real-time video four times faster on Trainium chips at half the cost of equivalent Nvidia GPU inference — a company-reported figure, not an independent benchmark, but one that represents the direction of improvement the technology is intended to deliver.

According to SiliconANGLE's reporting, Decart has raised more than $400 million in total funding and was valued at approximately $4 billion in a May 2026 funding round. The reported $7 billion acquisition price represents a 75% premium over that last private valuation — significant for a deal that has not yet closed and involves a startup with no disclosed revenue.

The Nvidia Dependency Problem Amazon Is Trying to Solve

To understand why Amazon would pay $7 billion for chip portability software, you need to understand the scale of the problem it is trying to solve.

Amazon Web Services spends billions of dollars annually on AI compute infrastructure. A significant portion of that spend flows to Nvidia in GPU purchases, and the AI workloads running on those GPUs are built on CUDA — Nvidia's proprietary programming framework that represents the deepest and most durable technology lock-in in the history of enterprise computing.

The depth of the problem is architectural. When you train a model on Nvidia GPUs using CUDA, the model's weights are optimized for the specific numerical precision and memory bandwidth characteristics of Nvidia hardware. Running that model on a different chip requires either accepting significant performance degradation or reoptimizing the model's architecture for the target hardware. This is months of engineering work per model, and most enterprises do not have the technical capacity to do it in-house.

The strategic consequence is that Amazon, Google, and Microsoft have each invested billions of dollars building custom AI chips — Trainium, TPUs, and Maia respectively — that are significantly cheaper to operate than Nvidia GPUs for inference workloads, but that are difficult to migrate existing AI workloads to because those workloads were built for CUDA.

Nvidia's $12.9 billion acquisition of Hugging Face earlier this year was partly about this same dynamic: controlling the open-source model hub that most developers use means Nvidia has influence over which chip ecosystems those models are optimized for. The hyperscalers are competing against Nvidia's distribution moat with infrastructure investments; Decart represents a different approach — solving the migration problem at the software layer rather than the distribution layer.

Why Amazon Values Decart More Than Anthropic Did

The reported sequence is strategically revealing. Anthropic, according to Bloomberg reporting cited by multiple outlets, held discussions with Decart about a roughly $6 billion acquisition earlier in 2026 and chose not to proceed after due diligence. Amazon is now in discussions at approximately $7 billion — a higher price for the same asset Anthropic evaluated and passed on.

The reason the same asset is worth more to Amazon than to Anthropic comes down to the difference in their positions relative to the chip dependency problem.

Anthropic is a model company. Its primary asset is the Claude model family. Chip portability matters to Anthropic primarily as an infrastructure cost reduction opportunity — if Anthropic could run its inference workloads on cheaper Trainium chips rather than Nvidia GPUs, it would reduce per-token compute costs. But the strategic value of chip portability is secondary to model quality for Anthropic. The $6 billion price was apparently too high relative to the cost reduction Decart's technology would provide to a model company.

Amazon is a cloud infrastructure company. Its primary business is selling compute. If Decart's software allows Amazon's customers — who have models trained on Nvidia CUDA — to migrate those workloads to Trainium with minimal engineering effort, Amazon does not merely save compute costs. It captures revenue from compute that currently flows to Nvidia rather than to AWS. The addressable market for Decart's value proposition is every AWS customer with a CUDA-trained AI inference workload. At AWS's scale, that represents tens of billions of dollars of annual AI compute spend that Amazon could redirect from Nvidia's hardware to its own.

This is the strategic logic that justifies a 75% premium over Decart's last private valuation: Amazon is not buying a startup. It is buying the conversion tool for a migration it has been trying to execute for four years.

The Chip Independence Race Among Hyperscalers

Amazon is not the only hyperscaler making infrastructure bets to break Nvidia's hold on the AI compute stack.

AMD's $8.2 billion acquisition of World Labs was similarly structured around reducing Nvidia's dominance — in AMD's case by acquiring spatial intelligence capabilities that made AMD chips more attractive for physical AI workloads, where AMD could compete on price and Nvidia's architectural advantages were less pronounced.

Google has taken a different approach: building the TPU ecosystem and investing heavily in JAX, an open-source machine learning framework that runs efficiently on TPUs without requiring CUDA. Google's strategy is to build an alternative development stack from the ground up rather than solving the CUDA migration problem with a portability layer.

Microsoft has Maia, its custom AI accelerator, but has not made a major software-layer acquisition to solve the migration problem — a potential strategic gap relative to Amazon and AMD's more aggressive M&A posture.

The competitive pattern is clear: every major hardware player is making M&A and infrastructure bets around the same strategic question — who controls the workload migration layer when enterprises want to move AI models between chips.

The $7 Billion Valuation Math

Decart has not disclosed revenue. This is the detail that makes the acquisition price difficult to evaluate on traditional metrics and that reportedly contributed to Anthropic's decision not to proceed.

The strategic premium interpretation is more likely to explain Amazon's number than a traditional revenue multiple. Amazon can calculate internally what the value of migrating X units of current Nvidia GPU inference to Trainium at Y percent lower cost would be — where X is knowable from AWS's own data on customer AI inference volumes, and Y is what Decart's technology claims to deliver. That calculation produces a value figure that may be independent of Decart's existing customer revenue, because the value creation happens within AWS's own infrastructure economics, not Decart's income statement.

At $7 billion for a company that raised $400 million and was last valued at $4 billion six months ago, the implicit assumption is that Decart's technology would generate sufficient compute cost savings or revenue capture to justify the price within a 5-7 year horizon. For a company spending tens of billions on AI compute annually, a technology that routes even a fraction of that spend toward cheaper proprietary chips rather than Nvidia GPUs can justify a $7 billion software acquisition.

The TypeSafe AI $870 million raise at a valuation built on strategic infrastructure positioning rather than current revenue illustrates the same pattern at smaller scale: infrastructure companies solving foundational AI compute problems are valued on strategic importance rather than current GAAP revenue, because the companies paying the highest prices are competing for position in an infrastructure market whose revenue will compound over the next decade.

The Enterprise AI Buyer Implications

For enterprise AI buyers, the Decart acquisition discussion has four concrete implications worth tracking now:

1. AWS Trainium becomes a more credible Nvidia alternative. If the deal closes and Amazon ships Decart's DOS technology as a managed AWS service, enterprise teams running CUDA-trained models will have a supported, AWS-managed migration path to Trainium. The current friction — Trainium is available but migration requires significant engineering effort — would reduce to a more automated process, making the economics of Trainium inference actionable for a much broader set of AWS customers.

2. Chip portability becomes a procurement requirement. If Decart demonstrates chip-agnostic AI infrastructure at enterprise scale, procurement teams will begin requiring portability guarantees from AI infrastructure vendors. The current situation — enterprise AI architectures de facto committed to Nvidia because migration costs are prohibitive — will start shifting as the tooling improves. The 81% of CIOs who already lack full oversight of their AI agent infrastructure will have chip vendor portability added to their governance requirements.

3. Nvidia's moat narrows on inference, not on training. Nvidia's CUDA lock-in is most durable for frontier model training, where the specific architectural advantages of H100 and B200 hardware are most pronounced. For inference — the ongoing operational cost of running trained models at production scale — alternative chips are already competitive on price, and Decart's technology specifically targets inference migration. Enterprises that distinguish training and inference workloads in their procurement strategy will find the most immediate value from Decart-style portability.

4. The M&A window for AI infrastructure companies got more expensive. Every startup that has built tooling reducing hyperscaler Nvidia dependency is now valued against a new comparable: a chip portability startup worth $7 billion to Amazon. This reference point will increase valuations for similar companies in the next 12-18 months as hyperscalers compete to acquire remaining portability tooling before a competitor does.

If the Deal Falls Through

The acquisition is not confirmed. Multiple sources note that the deal could still fall apart and that another bidder could emerge. If Anthropic chose not to proceed at $6 billion, there may be reasons beyond pure strategic calculus — technical due diligence findings about the technology's maturity in production enterprise environments, engineering team scalability concerns, or valuation disagreements that could resurface with Amazon as well.

If the deal falls through, Decart is not stranded. It has raised $400 million, has AWS as a major customer, and the public reporting has effectively established a $7 billion reference valuation — making additional capital straightforward to raise. The most likely alternative is that Decart raises at a $6-8 billion valuation and continues as an independent chip portability provider, potentially serving all three hyperscalers rather than being owned exclusively by one. An independent Decart that sells to AWS, Google Cloud, and Azure simultaneously may create more total value than one acquired exclusively by Amazon — a dynamic that makes the deal's strategic logic somewhat paradoxical at the negotiating table.

The Playbook for Infrastructure Startups in the Chip Independence Race

The Decart deal illustrates a specific class of startup opportunity that will continue to generate large acquisition offers over the next 24-36 months. The pattern: identify a foundational infrastructure dependency that hyperscalers and large enterprises are actively trying to escape, and build the software layer that makes escape possible. The most valuable startups in this pattern follow a consistent framework:

1. Solve the migration problem, not the build-from-scratch problem. The hundreds of billions of dollars of sunk cost in existing CUDA-based AI infrastructure represents a migration problem, not a new construction problem. Software that reduces migration friction — while maintaining performance parity — has a clearer enterprise value proposition than software that requires enterprises to rebuild their AI stack from scratch.

2. Target the largest recurring cost, not the largest one-time purchase. Training is a large one-time cost per model version. Inference is an ongoing recurring cost that scales with production usage. The addressable market for inference cost reduction is orders of magnitude larger than the market for training acceleration, and it compounds continuously as enterprise AI deployment grows.

3. Validate with the hyperscaler before seeking acquisition. Amazon was a Decart customer and gave Decart early access to Trainium before the acquisition discussions. This customer relationship validated the technology and simultaneously made Amazon the ideal acquirer. Strategic enterprise relationships with hyperscalers serve as both technical validation and natural acquisition funnel.

4. Build the engineering team as the moat. Decart's technology is valuable in part because the engineers who built it understand the architectural nuances of multiple chip ecosystems at a depth that takes years to develop. Acquiring companies in this category is acquiring the team as much as the technology. Retention packages, vesting cliff structures, and organizational design that keeps technical founders engaged post-acquisition are as important as the IP terms.

Takeaway: Amazon's reported $7 billion bid for Decart is not a bet on AI capabilities. It is a bet on chip independence. The Decart Optimization Stack is the migration tool Amazon needs to redirect tens of billions of dollars of annual AWS Nvidia GPU inference spend toward Trainium — its own chip that Amazon controls without paying Nvidia's hardware premium. The math behind that strategic value can justify a multiple over current revenue that no traditional financial model would produce. For enterprise AI buyers, the implication is clear: chip portability is about to become an infrastructure category with enterprise-grade managed services, and the window to negotiate Nvidia-exclusive contracts without portability provisions is closing faster than most procurement teams have noticed.