What Amazon's AI capex means for legal tech buyers

A legal AI renewal now needs one more map than most vendor packets provide. If the product depends on Amazon Bedrock, Claude, Trainium, or the stack underneath Amazon Quick, the buyer should know which upstream compute and model providers can touch the workload before the order form is signed. The useful questions are not limited to whether the application vendor has a SOC 2 report or a polished privacy page. They are whether the workload can be re-routed, where it can be processed, who can log it, what happens when a model is retired, and which contract gives the legal department a remedy when the upstream layer fails.

That is the practical meaning of Amazon’s AI capex cycle for legal tech buyers. The spend is not just an infrastructure headline. It is becoming part of the supply chain for legal research, contract review, litigation support, patent drafting, knowledge search, document analysis, and agentic workflows. A buyer may contract with a legal application vendor and still carry operational dependence on Amazon, Anthropic, and, in some Claude deployments, several other cloud providers it never negotiated with directly.

The capex numbers matter because the dependency has moved upstream

Amazon guided to roughly $200 billion of capital spending for 2026 on February 5, after $131.8 billion in 2025, and reported $44.2 billion of Q1 capital expenditure, up 77% year over year.[1] CNBC put the broader 2026 AI spending plans of the major cloud and platform companies near $700 billion, a useful backdrop for buyers trying to understand why compute access, accelerator supply, model availability, and cloud commitments are now procurement facts rather than distant market context.[2]

The Q2 frame is more delicate. The Register, in a July 31, 2026 analysis of Amazon’s Q2 earnings materials, reported that Amazon had raised the 2026 capital spending figure to $220 billion and that AWS backlog stood at $496 billion.[3] A legal procurement file should reconcile those items against the Amazon Q2 earnings release before treating them as settled company disclosures. For diligence purposes, however, the reported direction is already enough to change the question: not “is Amazon large enough to stay in AI?” but “which legal workflows are now exposed to Amazon’s allocation, pricing, availability, and product-control decisions?”

Scale cuts both ways. It reassures the buyer that Amazon is not treating AI infrastructure as a side project. It also means the legal department may be accepting concentration at a layer where it has little direct leverage. If a legal AI vendor says its system is “built on Claude” or “available through AWS,” that is not a diligence conclusion. It is the beginning of the dependency map.

The same point applies to sales claims that convert analyst commentary into procurement comfort. A statement that most of Amazon’s capex is AI-related should be treated as an estimate unless the vendor can point to a company disclosure that says so. Even if the estimate is directionally plausible, it does not answer the legal buyer’s contractual questions: where will the data go, who can access logs, what model version is committed, and what remedy exists if the service misses a matter deadline?

The Anthropic-AWS machine now sits beneath prominent legal AI tools

Amazon’s relationship with Anthropic is where the legal-tech implications become concrete. Anthropic announced an expanded collaboration with Amazon for up to 5 gigawatts of new compute and said more than 100,000 customers use Anthropic models on Amazon Bedrock.[4] TechCrunch reported that Amazon added a $5 billion investment on top of an earlier $8 billion commitment, while Anthropic pledged more than $100 billion in AWS spending over ten years.[5]

Those figures do not prove that any individual legal tool is safer, faster, or more accurate. They do show why the upstream stack deserves its own line in the risk register. A legal AI application that relies on Claude through Bedrock inherits more than a model capability. It inherits a commercial and operational relationship between Amazon and Anthropic that can shape capacity, model access, regions, pricing, feature release timing, and continuity.

Fortune reported in May 2026 that Harvey, Legora, Solve Intelligence, and Eve were among legal tools built on Claude, and that legal was Anthropic’s top Claude Cowork power-user function.[6] That is an adoption signal, not an effectiveness finding. It does not show that those tools eliminate hallucination risk, satisfy privilege obligations, or fit every legal workflow. It does show that Claude is not a peripheral model choice in the legal market. It is already part of the application layer that law firms and in-house teams are buying.

Amazon Quick belongs in the same diligence conversation, but with care. Artificial Lawyer described “Amazon Quick for Legal” in July 2026, while AWS’s current framing is “How legal teams use Amazon Quick.”[7][8] That wording matters. Buyers should not treat Quick as a dedicated legal vertical product merely because the use case page speaks to legal teams. The diligence question is still the same: what model, what region, what retention setting, what log policy, what service commitment, and what exit path?

Claude routing turns vendor diligence into coordination diligence

The easy procurement mistake is to ask the application vendor one set of AI questions and assume the answer covers the whole chain. Claude complicates that assumption. ComplexDiscovery described a five-provider routing picture involving AWS, Google Cloud, Microsoft Azure, Akamai, and xAI’s Colossus 1, and discussed Akamai’s reported $1.8 billion, seven-year Anthropic commitment as a legal-AI vendor-risk issue.[9]

The operational risk is not that multi-cloud routing is inherently improper. It may improve capacity, latency, or resilience. The risk is that the legal buyer’s questionnaire may not reveal which route is actually used for its matters, whether routing can change without notice, whether logs follow the same residency commitment as prompts and documents, and whether a failover path creates a different regulatory or privilege posture.

For a legal department, that distinction is not academic. A contract-review system may process merger schedules, employee investigations, board materials, source-code exhibits, health information, sanctions-screening notes, or privileged litigation strategy. If the buyer approved one region or one provider path, then later discovers that load-balancing, failover, or model availability shifted processing elsewhere, the remediation burden falls on the legal and security teams that approved the tool.

This is also where the familiar subprocessor list starts to look thin. A subprocessor schedule can name cloud providers without telling the buyer which provider handles which workload, which model endpoint is used, which logs are retained, whether training is excluded, how long telemetry is stored, or whether model outputs are replicated across regions. A legal AI product can pass a standard vendor intake and still leave the upstream route unresolved.

What the stack map should show before signature

The buyer does not need a diagram for its own sake. It needs a diagram that can be converted into contract language. At minimum, the vendor should identify the application layer, orchestration layer, model provider, model access path, cloud provider, region, logging location, retention period, and failover route for each material workflow. “Claude via Bedrock” is not the same answer as “Claude direct through Anthropic,” and neither answer is complete unless the buyer knows whether other routes can be used.

The route should be pinned where the legal or regulatory posture requires it. If the vendor insists on load-balanced or dynamic routing, the buyer needs the permitted routing universe in writing, including regions, providers, notice obligations, and a right to reject routes that break the buyer’s legal requirements. ComplexDiscovery’s questionnaire agenda for Anthropic-related legal AI risk emphasized the same kinds of controls: pinned routing versus load-balanced routing, retention and log export, breach-notification timelines, regional residency, model deprecation, and term flexibility through 2033.[9]

A useful internal procurement note can be shorter than a vendor’s AI white paper. It should say, for each workflow, what data enters the system, which model processes it, which provider hosts it, where logs and metadata sit, what contractual term controls the upstream dependency, and what event gives the buyer a termination or suspension right. If the vendor cannot answer, that uncertainty should be visible in the approval record rather than buried in a security questionnaire attachment.

The terms that need to move from questionnaire to contract

A questionnaire answer is not much help when a model is unavailable at 9 p.m. before a filing deadline. The relevant commitments need to survive into the order form, data processing addendum, security exhibit, SLA, or product-specific AI terms. Buyers should avoid leaving the most important promises in sales emails or procurement portals that are not incorporated into the agreement.

Routing disclosure and change control

The contract should identify the approved model providers, cloud providers, access paths, and processing regions. It should also say whether the vendor may change those elements unilaterally. If changes are allowed, the buyer needs advance notice, a description of the new route, a security and privacy impact summary, and a right to object or terminate where the change affects confidentiality, privilege, residency, regulatory obligations, or client commitments.

Residency that covers more than the uploaded document

Residency language often fails because it protects the obvious file and ignores the operational exhaust around it. Prompts, outputs, embeddings, metadata, trace logs, abuse-monitoring logs, administrator activity, backups, and support exports may matter as much as the source document. The buyer should require region commitments for each category or a clear statement that a category is not created or retained.

Retention, deletion, and log export

The contract should state retention periods for customer content, prompts, outputs, embeddings, and logs; whether any content is used for training or model improvement; and how deletion works at termination. For legal departments, log export is not a convenience feature. It is part of incident response, privilege investigation, audit defense, and internal accountability. A vendor that can inspect logs to defend itself should not leave the buyer unable to reconstruct what happened.

SLA remedies that match legal deadlines

Service credits rarely compensate for a missed filing, a delayed production, or a contract-review bottleneck during a signing. The SLA should distinguish ordinary support inconvenience from material workflow outage. For high-dependency deployments, the buyer should negotiate escalation windows, incident updates, workaround obligations, termination rights for repeated failures, and export assistance if the model or route becomes unavailable.

Continuity risk is not hypothetical in AI procurement. Buyers evaluating Amazon-linked AI products should keep a separate record of model shutdown and discontinuation terms, including lessons from prior Amazon AI product terms and shutdown scenarios. A related continuity analysis is available in What Legal Risks Does Amazon's AI Model Shutdown Create? and What Amazon Nova's 2025 Terms Said About Its Discontinuation.

Model deprecation and substitution

A legal team that validates a workflow on one model should not silently receive another. The agreement should require notice before deprecation, a migration period, access to prior outputs and logs, regression-testing support, and a right to reject substitutions that materially change performance, security, residency, confidentiality, privilege handling, or approved use cases. If the application vendor depends on Anthropic or Amazon for that notice, the buyer should know the length of the vendor’s own upstream notice period.

Breach notification and upstream incidents

Breach-notification timing should account for upstream providers. A vendor should not be able to delay notice to the buyer merely because it is waiting for a cloud or model provider to finish its own analysis. The contract should require prompt notice of suspected unauthorized access, misrouting, logging outside approved controls, loss of availability affecting material legal workflows, and any upstream incident that reasonably affects the buyer’s data or obligations.

Exit rights through the real dependency horizon

Long upstream commitments can outlast a legal department’s immediate subscription term. If the model ecosystem is being shaped by cloud commitments running many years ahead, the buyer should avoid being locked into a short-term application contract with weak exit mechanics. The agreement should include export formats, deletion certificates, transition assistance, termination rights for upstream route changes, and a clean way to suspend specific workflows without shutting down the entire platform.

How to treat Amazon Quick, Bedrock, and Claude in the approval record

The approval record should separate three things that vendors often blend together: the legal application, the model, and the cloud access path. A tool may present itself as a legal assistant, rely on Claude for generation or reasoning, and access that model through Bedrock. Another may rely on Claude directly. Another may use AWS services for storage or orchestration while using a different model path. Those are different risk profiles.

• For Bedrock-dependent tools, record the approved AWS regions, model IDs or model families, logging settings, retention settings, service quotas, support path, and whether the vendor can switch models inside Bedrock without buyer approval.
• For Claude-dependent tools, record whether access is through Bedrock, Anthropic directly, or another provider route; whether routing can include AWS, Google Cloud, Microsoft Azure, Akamai, or xAI infrastructure; and what happens if the route changes.
• For Amazon Quick use cases, record the current AWS legal-use framing, the actual services and models used, the customer’s administrative controls, and whether the deployment is covered by general AWS terms, product-specific terms, or separate enterprise commitments.
• For Trainium-linked claims, ask whether the reference is merely infrastructure positioning or whether it affects the buyer’s workload location, model availability, performance, pricing, or support commitments.

The record should also state what has not been verified. If the vendor cannot confirm failover routes, say so. If residency covers customer content but not logs, say so. If the vendor says it cannot provide model-version commitments because the upstream provider controls deprecation, say so. Procurement records are most useful when they preserve the unresolved dependency instead of converting it into a vague “AI risk accepted” line.

For teams building a broader verification layer around AI legal tools, the same discipline applies beyond Amazon and Anthropic. The practical habit is to test claims against records, keep the stack map current, and attach contract consequences to the parts of the workflow that matter. That verification approach is discussed further in How Lawyers Can Enter AI Legal Tech Through Verification.

The buyer still owns the verification duty

Amazon’s capex scale helps explain why AWS, Bedrock, Trainium, and Anthropic sit beneath so many legal AI conversations. It does not answer whether a particular legal department can use a particular tool for privileged documents, regulated data, litigation deadlines, or cross-border matters. Capacity is not a residency commitment. Cloud scale is not a breach-notification clause. A model partnership is not an SLA remedy.

Before procuring or renewing a legal AI tool built on Bedrock, Claude, Amazon Quick, or adjacent AWS infrastructure, the buyer should require routing disclosure, data-residency commitments, retention and log-export terms, breach-notification timelines, SLA remedies, model-deprecation protections, and exit flexibility through the relevant contract horizon. If the vendor cannot provide those answers, the legal team should treat the missing map as part of the risk being approved.

References

1. Amazon sees 50% boost to capital spending this year, shares tumble — Reuters, February 5, 2026
2. Tech AI spending approaches $700 billion in 2026, cash taking big hit — CNBC, February 6, 2026
3. Amazon's Q2 was great, but the earnings release is packed with baloney — The Register, July 31, 2026
4. Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute — Anthropic
5. Anthropic takes $5B from Amazon and pledges $100B in cloud spending in return — TechCrunch, April 20, 2026
6. Even as hallucinations show up in legal filings, Big Law goes all in on AI with new Anthropic release — Fortune, May 12, 2026
7. Meet Amazon Quick For Legal – Updated — Artificial Lawyer, July 15, 2026
8. How legal teams use Amazon Quick — AWS
9. What Akamai's reported Anthropic deal means for legal-AI vendor risk — ComplexDiscovery