The Model Is No Longer the Moat

For most of the last three years, enterprise AI strategy started with one question: which model? The answer shaped budgets, vendor relationships and architecture. In August 2026, Alibaba published the weights of Qwen 3.8 Max, a 2.4-trillion-parameter model, for anyone to download and run. A year earlier, a model of that class would have been a closely held product. Now it is a free input, and the model question carries far less weight in a planning cycle than it once did.

Models still matter. What has changed is how hard it has become to charge a premium for one. When a product commoditizes, its margins usually move to the layers around it, as they did when computing moved to the cloud. Benchmarks, price lists, earnings filings and enterprise deployments suggest the same shift is under way in AI. Lasting value is moving down the stack to whoever controls compute, and up the stack to whoever owns and governs the workflows that run on it.

Table 1 · The advantages that defined AI strategy are being competed away 

The former moat, what is eroding it and where advantage is moving

Figures are drawn from the sections that follow: the capability gap from Artificial Analysis (April 2025 to July 2026; see Figure 1); price declines from Epoch AI (median rate for a fixed level of performance); capacity from Amazon's July 2026 earnings call; the model swap from Alibaba Model Studio and DeepSeek documentation; and the underwriting example from Kurady et al., a September 2026 preprint not yet peer-reviewed.

Models are getting cheaper

The clearest sign of commoditization is that the best free models are now close behind the best paid ones. Artificial Analysis, an independent benchmarking firm, tracks a composite Intelligence Index across reasoning, coding and knowledge tests. In April 2025, the top open-weight model trailed the top proprietary model by about 13 points. By April 2026 the gap was 6. In July, Kimi K3 scored 57 on the index, against 61 for Claude Opus 5, the top proprietary model: a gap of 4 points. 

Figure 1 · The capability gap between open and closed models is closing

Gap between the highest-scoring open-weight and proprietary models, Artificial Analysis Intelligence Index

Prices have fallen faster still. Epoch AI, a research group that tracks AI trends, estimates that the cost of reaching a fixed level of benchmark performance falls by a median of about 50× a year, and closer to 200× a year for models released since 2024. Artificial Analysis found that 9 of the 13 models offering the best intelligence for their price are open-weight. The leading ones offer comparable intelligence at between one-half and one-sixth of the price of proprietary equivalents. DeepSeek's current Flash model lists at $0.15 per million input tokens off-peak.

Anyone building a budget on these numbers should keep two qualifications in mind. Proprietary models still lead clearly on the hardest reasoning and agentic coding tasks. And a lower price per token does not guarantee a lower bill. Agents make many model calls per task, and providers reprice as their own costs move. DeepSeek now charges double its off-peak rates during peak hours. For finance teams, a useful AI forecast starts from workload volume and calls per task, with the rate card as one input among several. The overall direction still holds: model access is increasingly priced on cost, the way commodity inputs are.

The money moves into physical infrastructure

If model access is being priced down, the capital is going into the physical layer, at a scale rarely seen. Amazon, Alphabet and Meta together spent about $175 billion on capital expenditure in 2024 and $295 billion in 2025. Their guidance for 2026 adds up to $545–570 billion, nearly double again. Microsoft is on a similar path: $115.9 billion in its fiscal year to June 2026, up 79%, with about $175 billion expected in calendar 2026.

Figure 2 · Hyperscaler capex has roughly tripled in two years

Capital expenditure by company, $ billions

NVIDIA's Data Center segment brought in $89.0 billion in the quarter to July 2026. That was up 117% from a year earlier and more than eight times the $10.3 billion it reported for the same quarter three years before.

Figure 3 · The hardware layer is capturing the spend

NVIDIA quarterly Data Center revenue, $ billions, Q1 FY24 (April 2023) to Q2 FY27 (July 2026)

The hardware layer is also getting more competitive. Buyers have more chip suppliers to choose from, which gives them more room to negotiate. In July 2026, AMD and Anthropic announced a partnership to deploy up to 2 gigawatts of AMD Instinct MI450-series GPUs, with AMD taking an equity stake of up to $5 billion. Google's TPUs and Amazon's Trainium chips give the two largest clouds their own alternatives. Enterprises and their providers now have more than one supply line for compute. For procurement teams, chip choice now belongs in the RFP alongside price and delivery dates.

Demand outpaces capacity

Spending at this scale would look like a bubble if customers were not committing to it, and they are. The best measure is contracted backlog: revenue customers have committed to that has not yet been delivered. At AWS, backlog grew from $195 billion in mid-2025 to $496 billion in mid-2026. Google Cloud's grew from $106 billion to $514 billion, including more than $50 billion added in the second quarter of 2026 alone. Alphabet says most of that backlog is standard Google Cloud contracts. It also includes sales of TPU systems installed in customers' own data centers, and Alphabet expects to recognize just over half of it as revenue within 24 months. Microsoft's commercial backlog rose 84%, to $678 billion.

Figure 4 · Contracted backlog has grown about 2.5× in a year

Contracted revenue not yet recognized, $ billions, end of Q2 2025 vs. end of Q2 2026

The providers say as much. On Amazon's July earnings call, Andy Jassy said that even at about $220 billion of capex, "we will still not have enough capacity to meet all the demand we have in 2026." Asked in July how Google allocates its TPUs, Sundar Pichai said the "first priority" is securing the compute needed to compete at the frontier, balanced against what its consumer products and cloud customers need.

That leaves an enterprise that rents all of its AI capacity in an uncomfortable position. In a shortage, providers serve their own priorities first, and customers without committed capacity wait their turn. AI capacity planning belongs in the same conversation as supply-chain risk, with the same questions about guaranteed allocation and fallback options.

Why some build their own

This helps explain a trend that stood out in 2026: large non-tech companies building their own AI infrastructure. Eli Lilly's LillyPod went live in February. It is an NVIDIA DGX SuperPOD with 1,016 Blackwell Ultra GPUs delivering more than 9,000 petaflops of AI performance. In March, Roche added 2,176 NVIDIA Blackwell GPUs, bringing its hybrid-cloud AI factory to more than 3,500 GPUs, which Roche calls the pharmaceutical industry's largest.

Judged on cost per token, these investments look hard to justify while API prices are falling. Lilly and Roche are buying the ability to run proprietary science continuously. NVIDIA and Lilly describe the difference plainly. A productive wet-lab team can physically make and test about 2,000 molecular ideas per target each year, while a computational "dry lab" can screen billions before anyone synthesizes one. The inputs are decades of proprietary research data that neither company would hand to a shared service lightly.

Cost and control are separate arguments, and most AI cost comparisons only make the first. A company can rationally pay more for compute it controls. Control buys capacity that no one else can claim, data that stays inside its own governance and predictable performance for work that runs in labs and plants.

Few enterprises should copy Lilly outright. Owning GPUs means keeping expensive hardware busy, hiring the people to run it and refreshing it every product cycle, and low utilization can erase any savings. For most organizations the realistic choices sit in between. These include reserved capacity with a specialist GPU cloud, a private or sovereign environment for sensitive workloads and routing that sends routine tasks to cheaper open-weight models while keeping frontier models for the hard ones. For most CIOs, the first practical step is a workload inventory that separates work that needs control from work that only needs a low price.

The options between renting and owning

Enterprises that will not build their own LillyPod still need someone to run the options in between, and a market has formed to do it. Specialist GPU clouds, often called NeoClouds, are growing faster than the hyperscalers. CoreWeave's revenue rose 112% to $2.58 billion in the second quarter of 2026, with a contracted backlog of about $104 billion. Nebius grew revenue 454% to $582 million, reaching $3.0 billion in annualized run-rate for its AI cloud. Alongside them, colocation operators are offering GPU-as-a-service, and enterprises are pooling GPUs across business units on shared, multi-tenant platforms.

Figure 5 · Specialist GPU clouds are growing faster than the hyperscalers

Year-on-year revenue growth, quarter ending June 2026

NeoClouds typically pitch compute that is 40–60% cheaper than the major cloud platforms, and in many cases the economics hold up. The gaps tend to sit in the services around the GPUs. As our analysis of the NeoCloud market puts it, "the compute is cheap, but everything around it is the problem." Many providers were built for training and lack production inference services. Some lack current SOC 2 Type II certification, high-performance storage or cost reporting a finance team can use. Vendor reviews should test those areas as carefully as the hourly rate, because they decide whether a low-cost cluster can carry production work.

Renting remains the sensible default for most workloads, provided it is planned. That means committed allocation for the work that cannot wait, providers vetted for enterprise readiness and an architecture that lets each workload move between providers. In practice, control requires a named owner accountable for where each workload runs and why, whether or not the enterprise owns any GPUs.

This is the kind of work WWT does with clients. We help decide where each workload should run across four paths: private data centers, colocation, NeoCloud or GPU-as-a-service and public cloud. We treat hybrid as the permanent state. In our AI Proving Ground, part of an Advanced Technology Center backed by more than $1 billion of infrastructure investment, clients test architectures and providers on real workloads before committing. We also help build the middle. With NVIDIA, WWT designed and delivered a proof of concept that helped Scott Data, a multi-tenant Tier III data center in Omaha, offer GPU-as-a-service to customers in healthcare, financial services and other regulated industries.

Anthropic's growth and the workflow layer

Compute is only half of the shift. The other half happens above the model, and the model labs themselves show it most clearly. If models are commoditizing, why has Anthropic's run-rate grown 65-fold in 19 months? By its own figures, its annualized revenue run-rate rose from about $1 billion at the start of 2025 to more than $5 billion by August 2025 and $14 billion by February 2026. It topped $30 billion in April, and by the end of July it had passed $65 billion, according to Bloomberg.

Figure 6 · Anthropic's revenue run-rate is compounding

Anthropic annualized revenue run-rate, $ billions

Most of its business use is automation. Anthropic's September 2025 Economic Index found that 77% of business use of its API followed automation patterns, with the model completing whole tasks on its own. Its fastest-growing product is Claude Code, an agentic coding tool. The research firm SemiAnalysis estimated in February 2026 that Claude Code was authoring about 4% of all public commits on GitHub.

Customers are paying for a tool embedded in the repository, the build pipeline and the issue tracker, which becomes part of how a team works. The switching cost comes from the process that has been rebuilt around the tool, even though the model underneath could be replaced. Anthropic's growth supports the argument: its fastest gains come from workflow products, backed by large compute commitments like the AMD deal.

Why governance lasts

The workflow layer has its own weakness, and competitors are already exploiting it. Alibaba's Model Studio documentation shows developers how to run Qwen 3.8 Max inside Claude Code by changing a handful of environment variables. DeepSeek publishes an Anthropic-compatible endpoint for the same purpose. One reason customers still pay for Claude: proprietary models keep their lead on the hardest agentic tasks. A product that is only a thin interface over a model API can lose its workload to a single configuration change. The parts of the workflow layer that last are deep integration with a company's systems of record and the governance that makes autonomous work safe to trust.

Recent research shows why step-level controls are not enough. In a September 2026 preprint, Kurady and colleagues describe what they call compositional policy violations. These are failures where every step in an agentic workflow passes its own check, yet the workflow as a whole breaks policy. In their underwriting example, a loss-history check passes at $210,000, under a $250,000 referral threshold. Later documents add $75,000 in claims, bringing the total to $285,000, and no step checks again.

Step-level guardrails cannot catch this, because the policies enterprises care most about, such as authority limits, cumulative thresholds and review requirements, apply to the whole execution. Enforcing them takes trace-level auditing: a full record of where each input came from, state tracked across the workflow and policy re-checked as the work runs. That capability is expensive to build, specific to each organization's rules and systems and painful to replace, which is why competitors cannot easily copy it. The research is early and not yet peer-reviewed, but Gartner expects over 40% of agentic AI projects to be canceled by 2027, citing inadequate risk controls among the causes. Risk and audit leaders should ask of any agent deployment whether its controls evaluate the whole run or only each step.

What to own, what to rent

For enterprise leaders, the practical split is to rent what is commoditizing, keep switching easy and invest in the layers where value is concentrating.

Table 2 · Rent what is commoditizing; own what competitors cannot copy

Recommended approach for each layer of the AI stack

"Own" does not require owning hardware. Reserved or hosted capacity gives control when allocation is committed. Evidence for each row appears in Table 1 and the sections above.

Looking ahead

Models will keep getting better and cheaper, and every enterprise will have access to them. As value moves down into compute and up into governed workflows, the decisions that set organizations apart are about ownership. They include the capacity critical work depends on, the data no one else has and workflows governed well enough to let agents act. Those decisions belong on the agenda for the next budget cycle and board review, each with a named owner. Organizations that make them deliberately will capture the value moving away from the model. Those that do not will keep renting it back.