Andy Jassy’s “multibillion-dollar” statement was a 2024 annualized revenue run rate, not proof that AWS booked several billion dollars of standalone generative-AI revenue in its first quarter. Amazon did not disclose a separately audited AI-revenue line in the cited material. Jassy’s argument was that five developments—SageMaker adoption, demand for Nvidia and AWS chips, renewed modernization, Amazon Q, and the large on-premises opportunity—were turning AI into a broad AWS consumption engine.
What Jassy actually said
The claim came from Andy Jassy’s discussion of Amazon’s first-quarter 2024 results, as reported by CRN. He said AWS had reached a “multibillion-dollar revenue run rate” from generative AI.
A run rate annualizes the pace of business at a particular moment. It is not the same as quarterly recognized revenue, annual GAAP revenue, bookings, profit, or remaining contract value. AWS reported approximately $25 billion in first-quarter 2024 revenue, but Amazon did not provide a separately audited generative-AI figure in the cited account. “AI sales” can also include infrastructure, model services, databases, storage, networking, security and applications rather than one product category.
The five trends at a glance
| Trend in Jassy’s thesis | How AWS can monetize it |
|---|---|
| SageMaker adoption | Training, deployment, inference, storage, data processing, monitoring and governance |
| Demand for Nvidia and AWS silicon | EC2 accelerator capacity, networking, storage and related services |
| Renewed modernization | Broader migration and managed-service consumption |
| Amazon Q | Assistant subscriptions and the AWS workloads supporting data and applications |
| On-premises workloads | Cloud infrastructure and AI-enabled application modernization |
1. SageMaker becomes a control plane for model builders
Jassy presented Amazon SageMaker as a way for customers to handle data preparation, experimentation, training, inference optimization and model management in one AWS environment. That matters commercially because a model project can consume far more than training compute: it may also require storage, data processing, deployment endpoints, observability, security controls and ongoing retraining.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
#1 Best Overall
Customer examples cited by Amazon and CRN included Perplexity AI reporting models trained 40% faster on SageMaker, Workday reporting 80% lower inference latency, and NatWest reducing time to value from 12–18 months to under seven months. These are customer- or Amazon-reported outcomes, not independent benchmarks. Results depend on the workload, model, baseline, architecture and implementation; the figures do not establish a universal improvement.
AWS describes SageMaker as the more configurable option when a team needs control over custom models, training infrastructure, deployment and monitoring. Its Bedrock-versus-SageMaker decision guide contrasts that approach with Bedrock’s simpler API-based model access.
2. Nvidia demand and AWS custom chips
AWS offers Nvidia-based instances for customers that want broad framework compatibility and established tooling. It is also promoting its own silicon: Trainium for training and Inferentia for inference. AWS says newer Trainium generations are designed to improve performance and price-performance for large models, claims documented in its infrastructure overview and therefore requiring vendor attribution.
Jassy’s 2024 comments that larger quantities of Trainium2 would arrive in late 2024 and early 2025 were time-bound forecasts, not a current availability guarantee.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsThe buyer’s trade-off
- Nvidia: mature software, familiar frameworks and broad portability.
- Trainium or Inferentia: potential infrastructure economics for compatible, high-scale workloads, but possible compiler, porting, testing and support costs.
- Total cost: instance price is only part of the calculation; include engineering migration, utilization, networking, storage and regional capacity.
Regardless of the accelerator, AWS can sell surrounding compute orchestration, data services, networking and storage. A claimed chip advantage should not be treated as an independently verified customer saving.
3. Modernization as an AI spending catalyst
Jassy said many companies had completed much of their pandemic-era cost optimization and were returning to modernization and new initiatives. In his account, generative AI supplied a new reason to upgrade data estates and applications while reusing AWS security, partner, networking, storage and operations tooling.
Rank #3
That is Amazon’s explanation of customer behavior, not proof that every company finished cutting costs. AI projects can initially raise cloud bills, and deployments may be delayed by security reviews, weak data quality, uncertain returns, governance requirements or accelerator shortages. Some organizations will modernize selectively or keep systems hybrid rather than move everything.
The commercial logic is nevertheless important: an AI application can pull through spending on databases, data pipelines, identity, monitoring, private connectivity and application hosting even when the model itself is accessed through an API.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 114. Amazon Q targets the application layer
Amazon Q was positioned in 2024 as an assistant for software development and internal company information. Described capabilities included code generation, explanation, testing, debugging, refactoring, Java modernization, planned .NET transformation and multi-step agents that prepare implementation plans and apply changes across files and test suites.
Those are Amazon product-positioning claims, not an independent finding that Q is the industry’s most capable assistant. Q can create a direct software-revenue stream through subscriptions or user-based pricing while encouraging AWS usage for enterprise data access, permissions, hosting, security and underlying models. Product capabilities and prices change, so buyers should verify the current details on the Amazon Q Developer page.
5. The remaining on-premises opportunity
Jassy estimated that 85% or more of global IT spending remained on premises. That is his estimate, not a universally accepted accounting measure; the answer changes depending on whether the denominator includes all hardware, software, personnel and telecommunications spending or only infrastructure services.
The figure supports AWS’s strategic case: AI is both a new workload and a reason to move data and applications into cloud environments. It does not mean 85% of workloads will move to AWS. Organizations retain systems on premises for latency, sovereignty, compliance, resilience or economics, and many use hybrid or multicloud designs. AI may also run in private data centers, at the edge or on specialized infrastructure.
Best Value
Where Bedrock fits
The original five-part framing underplayed Amazon’s three-layer stack: infrastructure and chips, platform services such as SageMaker and Bedrock, and applications such as Q. Bedrock provides managed access to foundation models and application-building features; its model catalog and prices change frequently.
Bedrock pricing varies by model, modality, region, tier and usage mode. Token pricing is not automatically cheaper than compute-based SageMaker deployment. At high volume, API charges can be substantial; conversely, dedicated SageMaker endpoints can waste money when they sit idle or are poorly optimized.
Choose the layer that matches the job
- Bedrock: API-first applications using existing foundation models with limited infrastructure management.
- SageMaker AI: custom-model development, fine-tuning, training, deployment and monitoring requiring deeper control.
- Both: Bedrock for general foundation-model use and SageMaker for specialized models or controlled hosting.
What the thesis proves—and what it does not
Supported by the argument
- AI demand can generate consumption across several AWS layers rather than one chatbot product.
- SageMaker, Bedrock, chips and Q give AWS multiple monetization paths.
- AI projects often require data, security and application modernization.
Still unproven from the cited claim
- The exact dollar amount of AWS generative-AI revenue.
- Whether the run rate was sustained or converted into durable profit.
- Whether AWS chips are cheaper or faster for a customer’s specific workload.
- Whether the 85% on-premises estimate predicts actual migration.
Infrastructure supply also limits how quickly demand becomes revenue: GPUs and accelerators require power, data-center capacity, cooling, networking and regional availability. Model-provider licensing, prices and availability can change as well. Enterprise adoption may stall over sensitive-data exposure, permission leakage, hallucinations, auditability, regulation, prompt injection or excessive agent permissions.
A practical buying framework
- Define the workload. Separate training, batch inference, real-time inference, retrieval and coding assistance.
- Choose managed access or control. Start with Bedrock for an API-led application; evaluate SageMaker when custom training, hosting or monitoring is central.
- Benchmark actual economics. Test representative batch sizes, sequence lengths, latency targets and utilization rather than relying on a chip headline.
- Price the whole system. Include data movement, storage, networking, security, observability and engineering labor.
- Set governance gates. Validate permissions, data residency, evaluation, audit logs, human review and agent scope before production.
- Keep deployment flexible. Hybrid or multicloud may be the right answer where compliance, latency, resilience or portability outweighs consolidation.
Jassy’s 2024 thesis remains a useful map of AWS’s commercial strategy, but its central number must stay in its historical context: a company-stated annualized pace, not a separately reported quarterly AI revenue total.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
