Twenty nine parts built the machine. This one grades it. AWS did not take the lead in generative AI by shipping the smartest model in any given month, and it rarely has. It took the lead by being the place a hundred thousand organizations already ran everything else, then giving them a way to add models without leaving.
So the honest verdict is not a benchmark score. It is a fit question. This closing part is my scorecard for the AWS generative AI stack after walking every layer of it, where it earns the default choice, where it quietly costs you, and the specific situations where I would send a team to Azure, Google Cloud, or IBM instead. No new services here, just judgment on the ones you already met.
Bottom line up front
AWS wins on breadth and on silicon. One API reaches Anthropic, Meta, Mistral, Cohere, NVIDIA Nemotron, and Amazon Nova, and Trainium and Inferentia give you a cost lever no rival fully matches. If you already live on AWS, Bedrock is the safe default.
AWS loses on simplicity. The stack is an assembly kit, not a paved road, and the friction shows up in quotas, region gaps, and a console that sprawls. You pay for breadth in complexity.
My call: pick AWS when you value optionality and are already here. Look elsewhere when you want the single best frontier model wired in, the cheapest training silicon at scale, or governance certified for a regulated industry out of the box.
How I grade a cloud GenAI platform
A verdict is only as good as the axes behind it, so here are mine. I score a cloud generative AI platform on five things, because these are what actually decide a build. Model breadth, meaning how many good models you can reach and switch between without re-plumbing. Silicon and cost, meaning whether the provider owns chips that move the price of training and serving. Agents, meaning how production ready the tooling is for systems that call tools and take actions. Governance, meaning security, data residency, evaluation, and audit that a serious enterprise needs. And simplicity, meaning how quickly a competent team gets from zero to a running feature without fighting the platform.
No cloud tops all five. That is the whole reason a verdict has to be about fit rather than a single winner. Azure trades breadth for a tight OpenAI and Microsoft story. Google Cloud trades some enterprise polish for research grade models and cheap custom silicon. IBM trades raw model choice for governance built for regulated shops. AWS makes its own trade, and naming it is the point of this part.
One caution before the scores. A platform verdict ages fast. Every number and product name here reflects the stack as of mid 2026, and the model catalog in particular turns over in weeks, not years. Treat the shape of the argument as durable and re-check the specifics against the live pricing and model pages before you commit a roadmap to them.
What AWS gets right
Start with the strength that carries the rest. Model breadth on Amazon Bedrock is real and it matters. Through one API you reach Claude from Anthropic, Llama from Meta, Mistral, Cohere, AI21, Amazon Nova and Titan, and as of March 2026 NVIDIA Nemotron 3 Super. Switching models is a parameter change, not a migration, which means a bad bet on a model this quarter is cheap to reverse next quarter. On a market where the best model keeps moving, the ability to move with it is worth more than being locked to whoever leads today. That optionality is the single most defensible thing about the platform.
Second, silicon. AWS designs Trainium for training and Inferentia for inference, and they give you a cost lever the model-only clouds cannot fully answer. Not every workload fits them, the Neuron software path has real edges, and I covered exactly when each wins in Part 7. But for high volume, steady workloads, owning the chip lets AWS price aggressively, and the Amazon Nova family rides that advantage. Nova Micro, Lite, and Pro land at roughly 75 percent below comparably capable models in their class, and the 2026 additions, Nova 2 Sonic for speech to speech with a million token context and Nova 2 Omni for unified multimodal, extend that price to performance story into new modalities.
Third, the enterprise plumbing is genuinely deep. PrivateLink and VPC endpoints keep model traffic off the public internet, KMS gives you control of the keys, IAM ties access to the same identity model as the rest of your account, and Bedrock Guardrails, model evaluation, and CloudWatch logging cover safety, quality, and audit. For a security team that already reasons in AWS terms, none of this is new vocabulary. That continuity is a quiet, large advantage that no benchmark captures, and it is the real reason existing AWS shops rarely leave.
Where the stack still fights you
Now the other side, because a scorecard that only praises is marketing. The first cost of all that breadth is complexity. AWS hands you parts, not a paved road, and assembling a production system means picking a model, a vector store, a guardrail config, an agent runtime, a logging path, and a deploy pipeline, each with its own console and its own sharp edges. A team that just wants a working assistant by Friday feels this. The paved-road clouds get you to a first feature faster, and pretending otherwise does no one a favor.
The second cost is operational friction. Model availability differs by Region, quotas throttle you before scale does, and cross-region inference exists precisely because the capacity is not uniform. I have watched a launch stall not on code but on a quota increase sitting in a queue. None of this is fatal, all of it is planning tax, and you should budget for it rather than discover it. The region and quota reality is laid out in Part 8.
The third is a moving target where you least want one, the agent layer. Bedrock Agents was the original path, AgentCore reached general availability in late 2025 as the production answer, and having two overlapping stories is confusing while it settles. It will settle. But if you are choosing today, you are picking on a surface that is still shifting under you, and that is a real cost for a team that needs a stable foundation now. Frontier model quality on the very hardest reasoning tasks is the last gap. Nova is excellent value, not the outright quality leader, and when a task truly needs the best model on earth you are reaching for Claude or a rival, which Bedrock can serve, but then you are paying frontier prices and the cost advantage narrows.
AWS against the other three clouds
Lined up against the field, each cloud has a clear center of gravity. Azure is the Microsoft and OpenAI story, the right call when your shop runs on Entra, Microsoft 365, and wants OpenAI models with enterprise agreements. Google Cloud leads on its own research models, Gemini and Gemma, and on Cloud TPUs, which give it a cost to performance edge on the workloads that fit them, and industry reporting through early 2026 put Google Cloud growth ahead of both rivals in percentage terms, though from a smaller base. IBM watsonx is the governed choice, built for regulated industries that need factsheets, monitoring, and EU AI Act readiness more than they need the widest model menu.
AWS sits in the middle of that field as the breadth and integration play. It does not have the single tightest productivity-suite tie-in that Azure has, nor the cheapest research silicon story that Google tells with TPUs, nor the out-of-the-box regulated-industry governance that IBM leads with. What it has is the most models under one roof, its own competitive chips, and the deepest existing footprint to build on. For the large population of teams already standing on AWS, that combination is usually decisive, and the switching cost to chase a marginal edge elsewhere rarely pays back.
| Cloud | Platform lens | Flagship models | Best fit |
|---|---|---|---|
| AWS | Broadest catalog, own silicon | Nova, Claude, Llama, Nemotron | Already on AWS, want optionality |
| Azure | OpenAI and Microsoft stack | GPT family via Foundry | Microsoft-first, wants OpenAI |
| Google Cloud | Research models, TPUs | Gemini, Gemma | Wants Google models, TPU economics |
| IBM watsonx | Governed, regulated industries | Granite | Compliance-first, hybrid or on-prem |
The four platform lenses side by side. Each has a distinct center of gravity. The platform-by-platform detail lives in the sibling series linked at the close.
A cost recap you can hand to finance
Cost is where the breadth pays off, if you use it. The lever that matters most is picking the right Nova tier for the traffic, because the price gap between tiers is enormous and most traffic does not need the top model. Here are the current on-demand rates, the numbers I actually estimate against. Default to the cheapest tier that answers your hard turns acceptably, and promote only the requests that fail it. Two other levers stack on top, batch inference at 50 percent off for anything nobody is waiting on, and prompt caching so you stop paying to re-read a fixed system prompt every turn. Both were covered in the pricing part, Part 6.
| Nova tier | Input per 1M | Output per 1M | Reach for it when |
|---|---|---|---|
| Nova Micro | 0.035 dollars | 0.14 dollars | Text only, latency critical |
| Nova Lite | 0.06 dollars | 0.24 dollars | Default multimodal assistant |
| Nova Pro | 0.80 dollars | 3.20 dollars | Harder reasoning, still cost aware |
| Nova Premier | 2.50 dollars | 12.50 dollars | Most complex reasoning on Nova |
Current on-demand Nova pricing per million tokens. Batch halves these. Confirm against the Bedrock pricing page before quoting finance, model rates change.
Worked example
An assistant handles 500 million input tokens and 150 million output tokens a month. On Nova Lite that is 500 times 0.06 plus 150 times 0.24, so 30 plus 36, which is 66 dollars a month. On Nova Pro the same traffic is 500 times 0.80 plus 150 times 3.20, so 400 plus 480, which is 880 dollars. On Nova Premier it is 500 times 2.50 plus 150 times 12.50, so 1,250 plus 1,875, which is 3,125 dollars.
Same workload, a 47 times spread from bottom to top tier. That gap is the entire cost argument. Sit on Lite by default, route only the turns that genuinely fail it up to Pro, and reserve Premier for the narrow set that needs it. Rates are current Nova on-demand pricing, re-check before you commit a budget.
Here is a small estimator you can run to sanity check a tier choice before a launch. It is a plain calculation, no API call and no charges, so you can paste it anywhere Python runs and adjust the volumes to your own traffic.
RATES = { # dollars per million tokens, on-demand
"nova-micro": (0.035, 0.14),
"nova-lite": (0.06, 0.24),
"nova-pro": (0.80, 3.20),
"nova-premier": (2.50, 12.50),
}
in_millions, out_millions = 500, 150 # tokens per month, in millions
for model, (rin, rout) in RATES.items():
monthly = in_millions * rin + out_millions * rout
print(f"{model:14} ${monthly:,.2f}/mo batch ${monthly/2:,.2f}/mo")Expected output: four lines, from nova-micro at 38.50 dollars a month up to nova-premier at 3,125.00 dollars, each with the halved batch figure beside it. Failure mode: hard-coding rates goes stale the next time AWS reprices, so treat the dictionary as a value you refresh from the pricing page, not a constant. A second trap is estimating on output tokens you guessed at, since output volume drives the bill far more than input on the higher tiers, so measure real output length before you trust the total.
When I would pick someone else
A verdict that never sends you elsewhere is not a verdict, so here are the cases where I would. If your organization runs on Microsoft, Entra for identity, Microsoft 365 for work, and you want OpenAI models under an enterprise agreement, Azure is the shorter path and the tighter fit. Fighting that current to stay on AWS rarely pays. The full Azure platform picture is in the Azure Gen AI Series.
If you want Google research models, Gemini and Gemma, or the TPU economics on training and serving that fit them, Google Cloud is the sharper tool, and its vector and grounding story is strong. That case is made in the Google Cloud Gen AI Series. And if you are in a regulated industry where model risk factsheets, monitoring, and EU AI Act readiness are the first requirement rather than an afterthought, IBM watsonx was built for that shape of problem, laid out in the IBM watsonx Gen AI Series. For the hardware layer under all of these clouds, the NVIDIA AI Series covers the GPUs most of them still run on.
Notice none of these is about AWS being bad. Each is a case where another cloud is a better fit for a specific constraint, Microsoft alignment, Google models, regulated governance. That is what a platform verdict looks like when it is honest. The vendor neutral version of how to choose, independent of any single cloud, is in the GenAI Series.
My take
After thirty parts, the thing I would tell a team choosing today is boring and true. If you already run on AWS, the burden of proof is on leaving, not on staying. The breadth, the silicon, and the integration usually outweigh a marginal model edge somewhere else, and the switching cost is always larger than the demo makes it look. Pick AWS for what it is, an optionality machine for shops already here, and stop shopping. If you are greenfield with no cloud allegiance, then, and only then, let the specific constraint pick the cloud.
Where AWS earns the default, and where it does not
So here is the verdict this series was building toward. AWS earns the default for any team already on AWS that values being able to move as the model market moves, and that has the engineering depth to assemble parts into a system. You get the widest model catalog through one API, your own cost-shaping silicon underneath, and enterprise plumbing your security team already speaks. That is a strong hand, and for the large population it describes, the right move is to build here and stop second guessing.
AWS does not earn the default when your first constraint is something another cloud owns, Microsoft and OpenAI alignment, Google models and TPU economics, or certified governance for a regulated industry. It also asks more of a small team than the paved-road clouds do, so if you need a working feature fast with a thin staff, weigh the friction honestly. Before you commit either way, validate three things: that you have priced the workload on the right Nova tier rather than the default one, that you know how much of your system sits in AWS specific services above the model call, and that the model you are betting on is reachable if today’s leader changes next quarter.
That is the series. If you started at Part 1 and followed through, you have the whole AWS generative AI stack in your head now, from the shared responsibility model to the reference architectures. Go pick one feature on your roadmap, name its pattern, price it on the right tier, and ship it. Then come back to the guide when the next one needs a different part.
References
- Amazon Bedrock, build generative AI applications and agents
- Amazon Nova foundation models
- Amazon Bedrock pricing
- Amazon Bedrock AgentCore
- Amazon Bedrock or Amazon SageMaker AI, decision guide
About The Author
DrJha