AWS AI Revenue Is Booming: Why Amazon’s Rising Capex Matters
AWS AI revenue crossed a $25 billion annual run rate in the second quarter of 2026 — the same quarter Amazon’s trailing-twelve-month free cash flow flipped to a negative $7.6 billion. That contradiction sits at the center of the most important cloud business on earth. Amazon Bedrock is signing customers faster than it did in its first two years combined, AWS Trainium chips are locking in multi-gigawatt commitments from both Anthropic and OpenAI, and AWS just posted 37% year-over-year growth — its strongest stretch in eighteen quarters.
The demand side looks almost too good. Hundreds of thousands of customers are already on Bedrock. Spend in a single quarter exceeded everything that came before it. Trainium is no longer an experiment; it is becoming the foundation for some of the largest AI capacity deals in the industry. On paper, this is the moment AWS has been building toward for years.
Yet the money is leaving faster than it arrives. Capital expenditure is running well ahead of operating cash flow. Amazon raised its full-year cash capex outlook to roughly $220 billion, driven in part by the rising cost of memory. Free cash flow has swung from a solid inflow to a multi-billion-dollar outflow in the space of twelve months.
The Numbers Nobody Disputes
Start with the boring, audited stuff, because everything else is interpretation.
For fiscal 2025, AWS generated $128.7 billion in segment revenue, up 20% year over year, with $45.6 billion of segment operating income — a roughly 35.4% operating margin. That alone makes AWS one of the most profitable large-scale businesses ever built. Amazon’s total company operating income in 2025 was $80.0 billion, meaning AWS produced well over half of it on less than a fifth of the revenue.
Then the second quarter of 2026 happened. AWS revenue reached $42.2 billion, up 36.7%, which annualizes to a $169 billion run rate. Segment operating income hit $16.6 billion, up from $10.2 billion a year earlier. Do the division: the margin moved from about 33.1% to about 39.3%, an expansion of roughly six percentage points in twelve months. In a business this large, that is not a rounding error. That is operating leverage arriving in a quarter.
Amazon’s CFO noted on the call that Q2 operating income included roughly $600 million of tariff refunds and roughly $600 million from energy-contract fair-value remeasurement, both landing primarily in AWS. Strip out $1.2 billion and the AWS margin is about 36.5% — still a clear improvement, but a less dramatic one. Our calculation, not the company’s. This is the kind of detail that separates a headline from an analysis.
Now the other side of the ledger. Trailing-twelve-month operating cash flow through June 30, 2026 was $161.4 billion, up 33%. Trailing-twelve-month purchases of property and equipment were $173.0 billion, up from $107.7 billion a year earlier — a 60.7% increase. Capital spending, in other words, consumed roughly 107% of operating cash flow. Free cash flow swung from an inflow of $18.2 billion to an outflow of $7.6 billion, a $25.8 billion reversal that Amazon attributed primarily to “investments in artificial intelligence.”
And it is accelerating. Q2 2026 capex alone was $54.2 billion, versus $32.2 billion in Q2 2025. On the earnings call, Andy Jassy said Amazon now expects approximately $220 billion in cash capital expenditure for 2026, up from a prior estimate of roughly $200 billion. The reason he gave was not demand. It was the rising cost of memory chips.
That is a $20 billion guidance increase caused by a component price. Remember that when anyone tells you software businesses don’t have supply chains.
AWS AI Infrastructure
Interactive graphics on revenue growth, capex, cash flow, debt, backlog, and value capture.
AWS revenue is accelerating, but Amazon is spending more on infrastructure than operations are currently generating in cash.
AWS growth and margin expansion
Q2 2025 vs. Q2 2026. Revenue for Q2 2025 is implied from the stated 36.7% growth rate.
Capex is outrunning operating cash flow
Trailing twelve months through June 30, 2026.
The capital build-out is accelerating
Quarterly capex and revised 2026 cash-capex guidance.
Quarterly capex ($B)
2026 cash-capex guidance ($B)
Long-term debt nearly doubled
December 31, 2025 vs. June 30, 2026.
Backlog provides visibility — not certainty
Backlog compared with the annualized Q2 AWS revenue run rate.
Supplier gross profit can exceed AWS quarterly revenue
Directional comparison. NVIDIA data-center gross profit is an illustrative proxy.
AWS is an exceptionally profitable cloud segment funding an industrial-scale AI build-out. The investment works only if utilization arrives before depreciation and debt weigh on cash generation.
What AWS Is Actually Selling in the AI Era
The temptation is to say AWS sells compute. It doesn’t, not really. It sells avoided decisions.
Amazon Bedrock is the clearest expression of this. It’s a managed service that puts foundation models from many providers behind AWS APIs, wrapped in AWS identity, logging, private networking, and billing. In the second quarter alone, Amazon said it added more than ten fully managed models including OpenAI’s GPT-5.6, Anthropic’s Claude Opus 5, Google DeepMind’s Gemma 4, and SpaceXAI’s Grok 4.3.
Notice what’s happening there. AWS is selling access to models built by companies that are, in several cases, direct competitors of Amazon’s own model efforts — and in Google’s case, a direct competitor of AWS itself. That is a deliberate refusal to bet on a single model winner. Bedrock’s economic proposition isn’t “our model is best.” It’s “you don’t have to be right about which model is best, and you don’t have to move your data to find out.”
The adoption disclosures support the strategy. Amazon said hundreds of thousands of customers now use Bedrock, that more customers were added in the last six months than in the first two years after launch, and that customers spent more in Q2 than in all prior quarters combined. Those are management-reported indicators, not audited segment figures, and Amazon does not break out Bedrock revenue separately. But “more spend this quarter than all previous quarters combined” is a specific enough claim that it would be reckless for a public company to invent.
AWS Trainium sells something different: cost per token. Trainium and its inference sibling Inferentia exist because AWS ran the arithmetic on merchant accelerator margins and decided it would rather own that layer. In December 2025, AWS made Trn3 UltraServers generally available on Trainium3, its first 3nm AI chip — 2.52 FP8 petaflops per chip, 144GB of HBM3e, 4.9 TB/s of memory bandwidth, and up to 144 chips per UltraServer. AWS claims up to 4.4x higher performance, 3.9x higher memory bandwidth, and 4x better performance per watt than Trn2 UltraServers, and specifically on Bedrock, “up to 3x faster performance than Trainium2 with over 5x higher output tokens per megawatt at similar latency per user.”
That last metric is the tell. Output tokens per megawatt is not a marketing number. It’s the unit economics of AI inference costs expressed in the only currency that ultimately constrains a data center: electricity. Whoever wins on tokens per megawatt wins the inference cost war, because power is the input you cannot buy your way around with a bigger check.
All of these are company claims, and they compare AWS hardware to AWS hardware. Independent, workload-level total-cost benchmarks of Trainium versus NVIDIA remain thin in the public record. That gap matters, and we’ll come back to it.
The Pricing Model: A 300x Spread Hiding in Plain Sight
Bedrock pricing is not one price. It varies by modality, model provider, specific model, region, and service tier. Text generation is metered by input and output tokens, embeddings by input tokens, images by generated image. AWS also offers batch inference on selected models at 50% below on-demand rates, plus provisioned, reserved, flex, and priority tiers.
The spread inside that catalog is enormous. Reported pricing compiled by third-party cost-management firms puts Amazon’s own Nova Micro at roughly $0.035 per million input tokens, with premium frontier models running as high as $10 per million input and $50 per million output tokens. That’s a range of roughly 286x on input and 357x on output within a single storefront. (Reported figures from secondary sources; AWS’s pricing page is dynamic and changes by model, tier, and region, so treat any specific number as a snapshot rather than a fixed price.)
For a buyer, the headline token rate is close to useless on its own. The real equation is model inference plus retrieval plus embeddings plus vector storage plus compute and networking plus data transfer plus observability plus support plus engineering hours plus whatever provisioned capacity sits idle overnight. AWS’s own pricing explainer walks through a hypothetical chatbot and demonstrates exactly this: knowledge-base embeddings and vector-store costs can quietly dominate the bill even when the model itself looks cheap.
There’s a subtler trap. A cheaper model that needs more tokens to reach the same answer, or that produces outputs requiring more human review, can be more expensive per accepted outcome than a pricier one. Cost per token is the metric vendors publish. Cost per accepted business outcome is the metric that determines whether your AI project has positive ROI. They are not the same metric, and confusing them is the single most common budgeting error in enterprise AI right now.
The same problem appears across generative AI products: published pricing can hide the real cost of compute, retries, idle capacity, and human review. Our analysis of AI video generation pricing and unit economics shows how credits, model routing, and approval rates can turn an apparently cheap AI product into a compute-heavy business. The lesson applies to Bedrock too: the relevant benchmark is cost per accepted outcome, not cost per token.
Where the Revenue Actually Comes From
AWS doesn’t publish a revenue breakdown by service, so anyone claiming to know Bedrock’s exact contribution is guessing. What we have are management-defined indicators: an AI business above a $25 billion annual run rate growing triple digits, and a chips business also above $25 billion growing triple digits.
Those two numbers deserve scrutiny. They are not GAAP segments, their scope isn’t fully defined, and they may overlap. The “chips business” figure appears to include Graviton, Amazon’s ARM-based general-purpose CPU line, which is a mature product — Amazon said Graviton is used by 98% of its top 1,000 EC2 customers, that Graviton5 reached general availability with up to 30–40% better price-performance than comparable instances, and that Graviton revenue commitments rose nearly 3x quarter over quarter. So “$25 billion chips run rate” is not the same as “$25 billion of Trainium.” Amazon has never claimed it was, but the framing invites the conflation.
The more durable revenue mechanism may be the least glamorous one: cross-sell. AI applications need storage, databases, vector search, logging, security, CPU compute, and networking. Management’s argument is that AI consumption pulls core AWS consumption, because post-training, reinforcement learning, agent tool use, and data services all run on ordinary infrastructure. The public record doesn’t quantify how much of AWS’s core growth is AI-caused versus AI-correlated. But the logic is consistent with how AWS has always made money: sell one primitive, get paid for eleven.
Then there’s the backlog. AWS reported $496 billion in backlog, growing triple digits year over year. Against a $169 billion annualized run rate, that’s roughly 2.9 years of current revenue under contract. It’s the single most reassuring number in the entire story — and also the least verifiable, because the public record does not disclose delivery schedules, take-or-pay provisions, pricing, or termination conditions. Backlog is a promise. Revenue is a receipt.
The Unit Economics AWS Won’t Give You
Here’s where honest analysis has to stop and admit its limits. Amazon does not disclose customer acquisition cost, lifetime value, churn, net revenue retention, Bedrock gross margin, Trainium gross margin, accelerator utilization, or internal transfer pricing. Any article presenting those figures for AWS is manufacturing them.
What we can do is read the proxies.
On acquisition cost, Amazon announced a $1 billion investment in “AWS Forward Deployed Engineering” — AI engineers embedded directly with customers to co-develop agentic AI solutions, with early customers including Cox Automotive, the NBA, the NFL, Ricoh, and Southwest Airlines. It also committed up to $1 billion in cloud credits to accelerate U.S. Intelligence Community migration. Those are, functionally, customer acquisition costs. When a cloud provider starts paying its own engineers to sit inside customer offices and write the customer’s AI application, the sales motion has become a services motion, and services don’t scale like software.
On retention, the best available signal is that Bedrock spend in Q2 exceeded all prior quarters combined, and that Graviton commitments tripled sequentially. Expansion revenue appears strong. Churn is unknown.
On capital efficiency, the math is brutally clear. Trailing-twelve-month depreciation and amortization was $75.2 billion, with Q2 alone at $20.0 billion — annualizing near $80 billion. One year of $220 billion capex, blended across servers with five-to-six-year lives and data centers monetized over thirty-plus years, plausibly adds somewhere in the range of $35–40 billion of annual depreciation once assets enter service. That’s our estimate, not a company figure, and the true number depends on asset mix and timing. But even at the low end, a single year of spending creates a recurring depreciation charge equal to more than half of AWS’s entire current annualized operating income.
That is the return hurdle. Not “will customers buy AI?” They already are. The question is whether they buy enough of it, at high enough utilization and stable enough pricing, to absorb an $80 billion depreciation base growing toward $120 billion.
Jassy’s defense on the call was structural: servers and networking equipment break even, on average, in a little less than three years, then keep producing for at least five to six; data centers can be monetized for more than thirty years. Data centers get built early, servers get bought closer to demand. It’s a coherent capital-cycle argument. It is also, at this point, a management assertion awaiting proof.
The Funding: Amazon Is Now Borrowing to Build
For most of its history, AWS funded itself. That changed in 2026.
Amazon’s long-term debt reached $128.9 billion as of June 30, 2026, versus $65.6 billion at year-end 2025. The cash flow statement shows $67.0 billion of long-term debt proceeds in the first six months of 2026 alone. Property and equipment, net, climbed to $446.0 billion from $357.0 billion in six months. Cash stood at $78.2 billion with $44.8 billion of marketable securities.
Amazon is not remotely liquidity-constrained. But the financing mix has shifted, and debt-funded infrastructure is a different animal than retained-earnings-funded infrastructure. It introduces interest expense, refinancing exposure, and a harder floor under the ROIC calculation.
Amazon is not the only Big Tech company converting AI demand into hard assets, debt, and depreciation. Our analysis of Meta’s AI spending and free-cash-flow story examines the same transition from a cash-generative digital business toward a capital-intensive infrastructure model. The comparison matters because the real test is not whether customers want AI; it is whether future utilization can earn back the infrastructure built today.
Meanwhile, Q2 net income of $62.6 billion included $53.4 billion of non-operating pre-tax other income, primarily from Amazon’s investments in Anthropic. The quarter’s deferred income tax line jumped $17.7 billion, largely reflecting that gain. Back out the after-tax effect and Amazon’s operations-driven net income was closer to $27 billion — still excellent, and roughly a third of the reported figure. (Our calculation.)
Those Anthropic marks are real accounting, and they represent real value creation. They are also non-cash, non-operating, and completely unrelated to whether Bedrock is profitable. Any argument that “Amazon’s AI business is minting money, look at net income” is analytically backwards.
Who Actually Captures the Value
This is the question that should keep AWS strategists awake.
NVIDIA reported second-quarter fiscal 2027 revenue of $96.2 billion, up 106%, with Data Center revenue of $89.0 billion, up 117%, at a 75% GAAP gross margin. Apply that company-level margin to the Data Center line as an approximation and NVIDIA’s quarterly gross profit from data center alone is on the order of $67 billion — comfortably more than AWS’s entire quarterly revenue of $42.2 billion.
Read that again. The supplier’s gross profit exceeds the customer’s total revenue.
That single comparison explains Trainium better than any AWS keynote. Every dollar AWS spends on merchant accelerators is a dollar where the accelerator vendor books the fat margin and AWS books the depreciation. Custom silicon isn’t a vanity project; it’s margin repatriation. Anthropic’s commitment of more than $100 billion over ten years to AWS technologies, securing up to 5 gigawatts of new capacity, is precisely the kind of anchor volume that makes an in-house chip program pencil out — roughly $10 billion a year on average, or about 6% of AWS’s current annualized revenue, contracted to the one workload type Trainium was designed for.
The counter-risk is equally clear. That $100 billion is concentrated in a small number of counterparties, and Amazon is simultaneously Anthropic’s investor, its cloud provider, and its chip supplier. Ownership, commercial dependence, and control are three different things, and they are unusually tangled here.
The Competition Is Not Cooperating
AWS is still the largest cloud provider on earth, and it is losing share while doing it.
Synergy Research data reported by CRN puts global enterprise spending on cloud infrastructure services at $143 billion in Q2 2026, up $43 billion year over year — 43% growth, the highest in eight years. Within that, AWS held 28% share, down from 30% a year earlier. Microsoft held 20%, flat. Google Cloud reached a record 15%, up from 13%. Synergy’s chief analyst noted that GenAI-specific cloud services are growing 165% year over year.
There’s no contradiction between 37% growth and share loss — the market grew 43%, so AWS grew fast and still captured less than its proportional share of incremental spend. But over multiple years, that arithmetic compounds against you.
Google Cloud is the immediate problem. Q2 2026 revenue of $24.8 billion, up 82%, with $8.8 billion of operating income — a 35.5% margin on a $99 billion run rate. Google has TPUs, Gemini, its own data stack, and enough internal AI demand to keep utilization high. Its reported margin isn’t perfectly comparable to AWS’s, since Alphabet allocates certain shared AI research and infrastructure costs outside the Cloud segment. But a business growing 82% at a 35% margin is not a distant third anymore.
Microsoft attacks from a different direction: distribution. Intelligent Cloud revenue was $39.3 billion, up 32%; Azure and other cloud services grew 43%; Azure surpassed $100 billion for the fiscal year. Microsoft’s advantage is that AI consumption can arrive bundled inside a Microsoft 365 or GitHub relationship that was signed years ago. AWS has to win the infrastructure decision on its merits. Microsoft sometimes doesn’t have to make the decision at all.
The Moat, Honestly Assessed
What stops a competitor from replicating this tomorrow? Several things, in descending order of durability.
Data gravity is the strongest. Moving petabytes and rewriting the surrounding application is expensive and risky, and models change faster than data architectures do. Enterprise governance is next: identity, compliance, audit trails, data residency, and procurement relationships change slowly, and Bedrock’s integration with IAM, CloudTrail, CloudWatch, and private networking gives AWS a genuine incumbency advantage. Scale in power and land is third, and increasingly the binding constraint industry-wide.
What is not much of a moat: model access. If Bedrock’s core promise is “many models, one API,” then model neutrality is by definition commoditized — any competitor can also resell models. That’s why AWS is pushing so hard into AgentCore, adding Payments so agents can transact autonomously, Web Search for grounding, and Harness for stitching agent infrastructure together. The bet is that the durable value migrates from model access to agent orchestration, policy, and memory.
Which introduces its own risk. AWS closed Bedrock Agents Classic to new customers on July 30, 2026, directing them to AgentCore instead. That’s a rational product decision and a real customer cost. When the recommended architecture changes inside three years, migration risk becomes a line item in the buyer’s TCO model, not a technical footnote.
What Could Go Right, What Could Go Wrong
The bull case is genuinely strong. Demand exceeds available capacity — Jassy said Amazon does not expect to have enough capacity to meet demand in 2026 and probably 2027, with significant demand already visible for 2028. Backlog is nearly three years of current revenue. Margins expanded six points year over year. Trainium3 is delivering more tokens per megawatt, and if custom silicon takes real share of AWS’s own AI serving, AWS captures the margin NVIDIA currently earns. Cheaper inference expands the addressable application set, which drives more consumption of storage, databases, and networking, which are AWS’s highest-margin legacy businesses. In that scenario, the 2026 capex looks like 2013’s fulfillment centers: a scary chart that turned out to be the moat.
The bear case doesn’t require demand to disappear. It only requires timing to slip. Data centers get built before servers arrive; servers get installed before utilization ramps; memory prices rise 10% more and take another $20 billion of capex with them; model efficiency improves faster than expected and the tokens you built for get served with half the silicon; Google and Microsoft price aggressively to buy share during the land grab; and a $120 billion depreciation base meets a decelerating revenue line. Margins compress not because AWS did anything wrong, but because the asset base was sized for a demand curve that arrived eighteen months late.
The base case, on current evidence, is somewhere in between: continued strong growth, gradual share erosion, margins that hold up better than skeptics expect but wobble as depreciation lands, and free cash flow that stays weak or negative until capex growth decelerates below revenue growth. Amazon’s Q3 2026 guidance — net sales of $197–202 billion and operating income of $22.5–26.5 billion — implies a sequential operating income step down from $27.5 billion, which is worth watching closely.
What the Numbers Really Say
Three things, none of them what the headlines say.
First, AWS AI revenue growth is real and verifiable at the segment level. A 37% growth rate on a $169 billion run rate with a 39% operating margin is an extraordinary result, and the six-point margin expansion — even adjusted down to about 36.5% for one-time items — is evidence that AI workloads are not structurally destroying cloud economics. That was a genuine open question two years ago. It is now substantially answered.
Second, profitability at the AI layer specifically remains undisclosed. We do not know Bedrock’s gross margin. We do not know whether Trainium is profitable after engineering, compiler development, Neuron support, and idle capacity. We do not know AWS’s accelerator utilization. Amazon has given us adoption metrics and run rates where it could have given us margins, and companies generally disclose the flattering number.
Third, the balance sheet has become the story. Capex at 107% of operating cash flow, negative free cash flow, long-term debt doubling in six months, and a $220 billion spending plan set partly by memory chip prices — this is a capital-intensive industrial business wearing a software company’s valuation. That’s not a criticism. It’s a category correction.
The Financial Verdict
AWS’s AI platform looks like a strong business inside an expensive bet. The revenue is audited. The growth is accelerating. The margins are expanding. The customers are real and named. The backlog is enormous. And the free cash flow is negative because the company is choosing to buy the next decade at 2026 prices.
The honest verdict is conditional attractiveness. AWS is economically compelling for customers who already have data and applications there, who need enterprise governance, who want model choice without betting on a single lab, and whose workloads can run on Trainium without a rewrite. It’s less compelling for portable, price-sensitive workloads already optimized elsewhere. For Amazon, the strategy is coherent and commercially validated — but validation of demand is not validation of return. Those are different tests, and only the first one has been graded.
Here’s the lesson that outlasts this particular quarter. In software, growth and profitability usually travel together, because the marginal cost of another customer rounds to zero. In AI infrastructure, they can move in opposite directions for years, because the marginal customer needs silicon, memory, floor space, and megawatts that must be bought before they’re used. AWS’s AI revenue is growing 37% a year. Its capital spending is growing 61%. Whether this ends as the best infrastructure investment of the decade or the most expensive lesson in capital timing depends on a single question that no press release can answer yet: does the demand show up on schedule?
Everything else is detail.
Financial Disclaimer: This article is provided for informational and educational purposes only. Financial figures cited reflect company disclosures and third-party reporting available as of late August 2026 and may change. Calculations labeled as estimates rest on stated assumptions and cannot be independently confirmed; publicly available information about AWS’s AI-specific unit economics is incomplete, and Amazon does not separately report Bedrock, Trainium, or Inferentia financials. Nothing here constitutes financial, investment, legal, tax, accounting, or professional advice, and nothing here is a recommendation to buy, sell, invest in, use, or avoid any company, security, technology, product, or service. Readers should conduct their own research and verify current figures before making financial or business decisions.
Sources
Amazon.com, Inc., “Amazon.com Announces Second Quarter Results,” July 30, 2026 — https://www.sec.gov/Archives/edgar/data/1018724/000101872426000024/amzn-20260630xex991.htm
Amazon.com, Inc., “Amazon.com Announces Fourth Quarter Results,” February 5, 2026 — https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Fourth-Quarter-Results/
Amazon Q2 2026 earnings call transcript coverage, Fortune — https://fortune.com/company/amazon-com/earnings/q2-2026/
Seeking Alpha, “Amazon outlines Q3 net sales of $197B–$202B while lifting 2026 cash capex to about $220B” — https://seekingalpha.com/news/4622393-amazon-outlines-q3-net-sales-of-197b-202b-while-lifting-2026-cash-capex-to-about-220b
Fierce Network, “Amazon’s AI infrastructure bill just got $20B bigger” — https://www.fierce-network.com/cloud/memory-drives-amazon-capex-another-20b-2026
AWS, “Announcing Amazon EC2 Trn3 UltraServers,” December 2, 2025 — https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/
AWS, “AI Accelerator — AWS Trainium” — https://aws.amazon.com/ai/machine-learning/trainium/
AWS, “Amazon Bedrock Pricing” — https://aws.amazon.com/bedrock/pricing/
AWS, “Amazon Bedrock now offers OpenAI models, Codex, and Managed Agents (Limited Preview),” April 2026 — https://aws.amazon.com/about-aws/whats-new/2026/04/bedrock-openai-models-codex-managed-agents/
OpenAI, “OpenAI models, Codex, and Managed Agents come to AWS” — https://openai.com/index/openai-on-aws/
Anthropic, “Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute” — https://www.anthropic.com/news/anthropic-amazon-compute
CRN, “Cloud Market Share Q2 2026: Google Gains Share As AWS Falls” (Synergy Research Group data) — https://www.crn.com/news/cloud/2026/cloud-market-share-q2-2026-google-gains-share-as-aws-falls
Alphabet Inc., “Alphabet Announces Second Quarter 2026 Results” — https://s206.q4cdn.com/479360582/files/doc_financials/2026/q2/2026q2-alphabet-earnings-release.pdf
NVIDIA, “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027” — https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
CloudZero, “Amazon Bedrock pricing in 2026: every model and hidden cost” — https://www.cloudzero.com/blog/amazon-bedrock-pricing/
Supplied research document: “Amazon AWS + AI: A Forensic Public-Record Audit of Bedrock, Trainium, Inferentia, and the AI Inference Cost War,” research cutoff August 27, 2026
Comments
Loading comments...