AWS funds generative-AI proofs of concept on Amazon Bedrock with a dedicated, partner-filed credit pool — typically $10K–$50K. It is not on the public Activate page, it is not the same thing as your Activate balance, and it is approved against a use case, not a funding stage. This is the definitive guide: what counts as a fundable POC, the dollar bands, eligibility, the exact application mechanic, the timeline, and how to turn a funded POC into a production workload AWS keeps backing.
There are three different things people call "AWS AI credits," and conflating them is the single most common reason a Bedrock POC application stalls. The POC pool is the third, least-documented one — and the one purpose-built to fund a generative-AI experiment.
The first thing people mean is the Activate balance: the $1K (Builders), $5K (Founders), or $100K (Portfolio) general-purpose credit pool that covers any AWS service, including Bedrock inference. The second is the Generative AI Accelerator — a competitive, cohort-based program that awards large packages (up to $1M for a small number of startups globally) plus mentorship and go-to-market. The third — the subject of this guide — is the Bedrock proof-of-concept credit pool: a targeted, use-case-scoped grant that exists specifically to de-risk a single GenAI experiment on Bedrock.
The POC pool is not on the public Activate page. There is no "apply for Bedrock POC funding" button. It is provisioned through AWS partners and partner-development managers as a funding source attached to a specific opportunity record. In practice it behaves like a service-specific incentive: AWS wants teams to actually try Bedrock — to stand up a real workload against Claude, Llama, Titan, Nova, or another hosted model — and it is willing to fund the inference, storage, and surrounding services for the duration of a bounded experiment.
Why does AWS run a program like this? Because the gap between "interested in GenAI" and "running GenAI in production on AWS" is where most teams stall. Inference cost uncertainty, model-selection paralysis, and the fear of a runaway bill all keep teams in slide decks instead of in the console. A POC credit removes the cost variable for 60–120 days so the team can answer the only question that matters: does this work well enough to ship, and what does it cost at scale? AWS is betting that a meaningful fraction of funded POCs convert to ongoing Bedrock spend — and the data supports that bet, which is why the program persists.
The mental model that matters: the POC pool is approved against a workload, not against you. Activate Portfolio asks "is this a fundable company?" The Bedrock POC pool asks "is this a fundable experiment?" That distinction shapes everything downstream — who qualifies, what the application contains, and why a well-scoped bootstrapped team can win POC credits that a better-funded but vaguer team cannot.
Bedrock POC credits are a use-case-scoped, partner-filed credit grant (typically $10K–$50K) that funds a bounded generative-AI proof of concept on Amazon Bedrock — separate from your Activate balance, approved against the merits of the experiment rather than your funding stage.
The headline band is $10K–$50K. Where a specific POC lands inside that band is not random; it tracks the projected inference footprint of the experiment and the credibility of the path to production. Here is what each tier typically looks like.
The award is sized to the POC, so the reviewer is implicitly estimating "what will this experiment consume on Bedrock over its funded window, plus the supporting services around it?" A retrieval-augmented chatbot answering a few thousand queries a day on a mid-tier model has a very different inference profile than a document-processing pipeline running batch inference over millions of pages, and the credit reflects that.
Three things move the number up: a larger projected inference footprint (higher token volume, larger or more expensive models, more concurrent workloads), a credible and sizeable production trajectory (a reviewer funding a POC that plausibly becomes $10K+/month of Bedrock spend will fund more generously than one that tops out at $300/month), and supporting-service breadth (a POC that also exercises OpenSearch for vector search, S3 for a knowledge corpus, Lambda for orchestration, and Step Functions for pipelines reads as more substantial than a single API call).
Two things move it down or trigger a downgrade: an implausibly large projection relative to the team and use case (claiming millions of daily requests as a three-person team with no traffic invites scrutiny, not a bigger award), and vagueness — a POC with no defined success metric and no named model reads as exploratory rather than fundable, and reviewers cap exploratory requests at the low end of the band or decline them.
Typical shape: one well-defined GenAI feature — a support-ticket summarizer, a content classifier, a single RAG assistant over an existing knowledge base — on one model, with a clear two-to-three-month evaluation window.
Inference profile: modest. Hundreds to low thousands of requests per day, often on a cost-efficient model tier (e.g., a Haiku-class or Nova-class model) or light usage of a mid-tier model.
Who lands here: bootstrapped teams, agencies running a POC for one client, and ISVs validating a single feature before committing roadmap to it. The $10K tier is the most accessible and the fastest to approve.
Typical shape: a GenAI workload with more than one moving part — retrieval plus generation plus an evaluation harness, or two related features sharing infrastructure — with an explicit intent to move to production if metrics clear a defined bar.
Inference profile: meaningful. Mid-tier models (e.g., a Sonnet-class model) at a few thousand requests per day, or a smaller model at higher volume, plus vector search and orchestration services.
Who lands here: funded startups validating a core AI feature, teams with existing AWS traffic that lends credibility to the production projection, and ISVs embedding GenAI into a shipping product.
Typical shape: a larger experiment — an agentic workflow, a high-volume document or media pipeline, or a multi-model architecture — where the production version is plausibly a five-figure monthly Bedrock commitment and the team has the capacity to ship it.
Inference profile: heavy. Frontier or large models, high token volume, batch inference, or multiple concurrent workloads, with substantial supporting services (knowledge bases, agents, guardrails, observability).
Who lands here: teams whose POC is essentially a pre-production pilot. The $50K tier is the top of the standard POC band; beyond it, the conversation shifts to the Generative AI Accelerator or a custom arrangement with an AWS account team.
$50K is the practical ceiling for the standard Bedrock POC pool. If your GenAI ambition is structurally larger — a product whose entire thesis is generative AI, with a credit need in the six figures — the right instrument is the Generative AI Accelerator (cohort-based, up to $1M, 60–90 day cycle) or a negotiated package via your AWS account team, not the POC pool. Many teams do both in sequence: a fast POC credit now to validate, then a larger program once the POC proves the workload.
The defining feature of the Bedrock POC pool is that it decouples credit access from funding stage. You do not need a Series A. You do not need a tier-1 VC. You need a real, fundable generative-AI experiment and a partner willing to attest to it. That widens the eligible population considerably.
Activate Portfolio is gated by the "institutional vouch" — VC backing or partner attestation tied to your company's fundability. The Bedrock POC pool is gated by the experiment's merit. This is why teams that cannot access the $100K Portfolio tier can still unlock POC credits: a bootstrapped two-person team with a sharp, well-scoped RAG POC over a real document corpus is more fundable, for POC purposes, than a well-funded startup whose AI plan is "we'll figure out where to add AI later."
Across the situations CloudRoute partners file most often, four profiles qualify cleanly and one structurally does not. The qualifying profiles share one trait: a specific GenAI workload they can describe in concrete terms. The disqualifying pattern is the absence of one.
Two secondary eligibility factors matter at the margin. First, region and export controls: AWS credit programs follow US export-control rules, so teams in non-sanctioned regions across the US, UK, EU, MENA, India, and APAC are fine, while sanctioned jurisdictions are not. Second, the competitor filter: AWS will not fund a POC whose output competes with an AWS-native service — for example, a product that is itself a hosted-model-inference platform substituting for Bedrock. This filter is narrow and catches very few legitimate GenAI applications, but it is worth knowing the boundary exists.
This is the section that determines outcomes. A fundable POC is not "more ambitious" than a non-fundable one — it is more specific. Reviewers are pattern-matching for a bounded experiment they can reason about. Five elements, present together, make a POC fundable.
Think of the POC description as the unit of evaluation. The reviewer is not assessing your company; they are assessing whether this experiment is a real, scoped thing AWS should fund. The strongest applications read like a one-page experiment design, not a pitch. The five elements below are what turn a paragraph from "exploratory" into "fundable."
Not "add AI to our app." Instead: "support agents spend ~6 minutes per ticket manually summarizing history; we want a Bedrock-generated summary to cut that to under 1 minute across ~2,000 tickets/day." A concrete problem with a number attached signals the POC is grounded in something real.
Name the model (or the candidate set you will evaluate — e.g., Claude Sonnet vs. a Nova or Llama tier), and name the surrounding AWS services: a knowledge base and vector store for retrieval, S3 for the corpus, Lambda or Step Functions for orchestration, guardrails for safety, CloudWatch for observability. A POC that exercises Bedrock plus a few native services reads as a genuine AWS workload, which is exactly what the funding is meant to seed.
State how you will decide whether the POC worked: an evaluation metric (answer accuracy, summary quality, deflection rate, latency target), a cost-per-request ceiling, and a go/no-go threshold. "We will ship to production if summary quality clears 85% on our rubric and cost-per-ticket stays under $0.02" is a fundable success definition. "We'll see how it goes" is not.
A POC has an end. Two-to-four months is the typical funded window. Bound the scope to one workload, one corpus, one set of users. Reviewers fund experiments, not open-ended R&D — a defined start, milestones, and an evaluation date make the request fundable.
The reviewer's implicit return is conversion: a POC that becomes ongoing Bedrock spend. So state, honestly, what production looks like if the POC succeeds — projected monthly request volume, the model you would run, and the resulting Bedrock spend. This is the single biggest lever on the dollar amount, because it tells the reviewer what they are buying.
Before applying, write the POC as one paragraph and check it answers five questions: What problem? Which Bedrock model and services? How will you measure success? Over what bounded window? What does production spend look like if it works? If all five are answered concretely, you have a fundable POC. If any is hand-waved, fix it before filing — that gap is exactly what a reviewer downgrades.
There is no public form for Bedrock POC funding. The credit is requested as a funding source attached to a partner-filed opportunity record in AWS's partner systems. Understanding the mechanic explains why the partner you work with — and how well they file — directly affects the outcome.
AWS partners register customer opportunities through ACE (APN Customer Engagements), the gated portal in the AWS Partner Network. A partner with Advanced or Premier tier has ACE submission rights; some Select-tier partners do not have full rights, which is why partner tier matters. When a partner files an opportunity for your POC, they create a structured record and request the POC pool as the funding source on it.
The record is where your one-paragraph experiment becomes a formal request. The partner translates your inputs into the fields AWS reviewers read:
An AWS partner-development manager reviews the record. Because the POC pool is use-case-scoped, the review centers on the experiment's merit and the production trajectory rather than your cap table. Two filing details disproportionately affect approval speed and amount: whether the projected consumption is realistic and itemized (a credible, service-level projection approves faster and larger than a single round number), and whether the partner has a track record of filing POCs that convert to production spend — high-conversion partners are effectively pre-trusted, which is why routing to an experienced partner matters more here than for a vanilla Activate request.
The Bedrock POC pool does not replace your Activate balance — it adds to it, provided the POC is a genuinely distinct workload. Getting the stacking right is the difference between an additive credit and a silent downgrade to $0.
The governing rule across all AWS credit pools is no double-counting. A reviewer will not fund $100K of Activate Portfolio for "general AWS infrastructure" and then $25K of Bedrock POC for the same undifferentiated infrastructure — that is one workload described twice. They will fund Portfolio for your general platform and a Bedrock POC credit for a clearly separate generative-AI experiment, because those are two distinct workloads with two distinct consumption profiles.
In practice the clean stack looks like this: Activate Founders or Portfolio covers your baseline compute, database, and networking; the Bedrock POC credit sits on top, scoped exclusively to the GenAI experiment — its inference, its vector store, its orchestration. The two records describe different things, so both can be approved. Where teams get burned is filing a POC request that simply re-describes their general infrastructure with the word "AI" added; the reviewer approves the primary pool and zeroes the POC line.
For an AI-first team, the sequencing usually runs: take the Activate credit appropriate to your stage first (it is broad and faster), then file the Bedrock POC credit for the specific experiment. If the POC then proves a large production workload, that is the moment to consider the Generative AI Accelerator or an account-team package — but those are mutually exclusive with the POC pool for the same workload, not additive on top of it. The honest framing throughout: additive credits require additive workloads. If there is genuinely only one workload, take the single best-fit pool and stop — extra filings on one workload do not increase the total, they just create rejections.
| Instrument | Typical amount | Scoped to | Submission path | Wall-clock | Stacks with POC? |
|---|---|---|---|---|---|
| Bedrock POC pool | $10K–$50K | One GenAI experiment | Partner via ACE | 10–21 days | — |
| Activate Founders | $1K–$5K | General AWS usage | Self-serve / partner | 2–10 days | Yes (distinct workloads) |
| Activate Portfolio | $100K | General infrastructure | Partner via ACE / VC | 11–18 days | Yes (distinct workloads) |
| Generative AI Accelerator | up to $1M | AI-first company | Competitive cohort | 60–90 days | Mutex (same workload) |
| MAP (migration) | $25K–$500K+ | Cloud/on-prem migration | Partner via APN | weeks–months | Yes (different program) |
From the team's side, unlocking a Bedrock POC credit is a few hours of work spread across a two-to-three-week wall-clock. The long pole is not your effort — it is AWS's review queue. Here is the realistic sequence.
The single highest-leverage thing you do is the prep in the first day: arriving with the five fundability elements already articulated turns a multi-call back-and-forth into a single working session. Teams that show up with a defined problem, a candidate model, and a rough consumption estimate get filed faster and approved larger than teams that treat the discovery call as the moment to start figuring out their POC.
You describe the POC in a few sentences: the business problem, the Bedrock workload you have in mind, and your rough sense of scale. This is enough to confirm fit and route to a partner whose track record matches your use case (RAG, agents, document pipelines, etc.) and region. Founder time: minutes.
A 30–45 minute working session with the partner to firm up the five fundability elements: lock the named problem, choose the model or candidate set, sketch the Bedrock-native architecture, set the success metrics and evaluation window, and estimate the production trajectory. The partner converts this into the consumption projection. Founder time: about an hour, plus a short follow-up to confirm AWS account details.
The partner submits the opportunity through ACE with the Bedrock POC pool requested as the funding source, the itemized consumption projection attached, and the success metrics and production trajectory documented. Your only input here is providing your AWS account ID and a one-paragraph use-case confirmation. Founder time: minutes.
An AWS partner-development manager reviews the record against the POC's merit and the production trajectory. Well-scoped, realistically-projected POCs filed by high-conversion partners move fastest. On approval, the credits appear in your AWS billing console with a defined validity window, and you receive an AWS confirmation email. Total founder time across the whole process: typically two to four hours.
Three things compress the timeline: (1) arriving with the five fundability elements already defined; (2) a realistic, itemized consumption projection rather than a single round number; (3) a partner with a track record of POCs that convert to production spend. The reverse — a vague use case, an implausible projection, an inexperienced filer — is what turns a two-week approval into a month of back-and-forth or a downgrade.
The POC credit is the beginning of the relationship, not the end of it. The program exists because a meaningful share of funded POCs convert to ongoing Bedrock spend — and a POC run with conversion in mind both produces better evidence and positions you for the next, larger funding instrument.
Run the POC against the success metrics you committed to. The output you want at the end is not just "it worked," but a concrete artifact: measured answer/output quality against your rubric, observed cost-per-request, latency profile, and an extrapolated monthly Bedrock spend at production volume. That artifact is what justifies shipping internally, and it is also exactly the evidence that supports the next funding conversation.
Manage the inference economics during the POC so production is not a cost surprise. The levers that matter on Bedrock are familiar: choosing the right model tier for each task rather than defaulting to the largest model, using prompt caching where the workload has stable context, batching where latency allows, and right-sizing retrieval so you are not paying to stuff oversized context into every call. A POC that demonstrates a disciplined cost-per-request is far more likely to convert internally than one that proves the feature works but at an unviable unit cost.
When the POC clears its bar, the credit typically runs down as the workload goes live, and ongoing usage becomes paid Bedrock spend — which is the program working as intended. At that inflection point, if the production workload is large, you have earned the credibility to pursue the next instrument: a larger package via your AWS account team, the Generative AI Accelerator if you are AI-first, or cost-optimization and committed-use arrangements as steady-state spend grows. A funded POC that converted is the strongest possible input to any of those conversations, because you are no longer projecting a workload — you are running one.
The gap between an approved POC credit and a declined one is rarely ambition — it is specificity. The same idea, described two ways, produces two different outcomes. This is the contrast reviewers pattern-match on.
| Element | Non-fundable framing | Fundable framing |
|---|---|---|
| Business problem | "We want to add AI to our product." | "Cut agent ticket-handling time from 6 min to under 1 min across ~2,000 tickets/day." |
| Model + architecture | "We'll use an LLM." | "Claude Sonnet on Bedrock + a knowledge base over our docs in S3, orchestrated with Lambda." |
| Success metric | "See if it's useful." | "Ship if summary quality ≥85% on our rubric and cost-per-ticket <$0.02." |
| Scope + timeline | Open-ended exploration. | One workload, one corpus, 10-week evaluation window with milestones. |
| Production trajectory | Unstated. | "~60K requests/month in production → ~$X/month Bedrock spend if metrics clear." |
| Likely reviewer action | Decline or cap at the $10K floor. | Approve, sized to the projected footprint ($25K–$50K range). |
Situation: Wanted to validate a Bedrock-powered support assistant — RAG over their help center plus ticket summarization — before committing roadmap to it. The team had a clear problem (agents spending ~6 minutes per ticket on history) but no AI infrastructure and real anxiety about inference cost running away during the experiment. Not VC-heavy enough to lead with a Portfolio conversation, and unsure the POC pool was even open to them.
What CloudRoute did: Routed within 20 hours to a partner with a RAG-on-Bedrock track record. In a 40-minute working session they locked the five fundability elements — named problem, Claude Sonnet + a knowledge base over the help-center corpus in S3, success metrics (summary quality ≥85%, cost-per-ticket <$0.02, deflection lift), a 10-week window, and a projected ~60K-requests/month production trajectory. The partner filed the ACE record requesting the Bedrock POC pool on day 4, with an itemized inference + OpenSearch + Lambda projection. A small Activate Founders balance covered baseline infra separately so the two records described distinct workloads.
Outcome: POC credit approved at $25K within 13 days, stacked on top of the existing Activate balance. The POC ran the full window, cleared its quality bar, and landed cost-per-ticket under target; the team shipped to production and the workload converted to ongoing paid Bedrock spend. CloudRoute's commission was paid by the partner from AWS's engagement funding — the customer paid $0.
POC window: 10 weeks · founder time: ~3.5 hours · credits secured: $25K · cost to customer: $0
CloudRoute routes you to a vetted AWS partner who scopes the POC with you and files the Bedrock POC funding request via ACE. Customer pays $0 — AWS funds the pool and the engagement. No procurement, no discovery theater.