The whole pitch fits in one sentence: the data never leaves

Here's the deal. On August 3, New York-based Superblocks and AWS announced a multi-year collaboration. On the surface it reads like any other partnership release — marketplace listing, co-selling, field enablement, joint workshops. But the center of gravity isn't distribution. It's where the software runs. The Superblocks platform itself deploys fully managed inside the customer's own AWS account, inside the customer's own VPC.

The same day, Superblocks published its 3.0 launch post with a title that leaves nothing to interpretation: secure private vibe coding on AWS, built for the age of AI cyberattacks. That's marketing language, but it tells you exactly where the company is aiming. Generating apps from prompts stopped being novel a while ago. What's new is that those prompts now touch production data — and the moment that happens, the product being sold shifts from speed to control.

Superblocks CEO Brad Menezes gave TechCrunch the one-line version: "We're going to bring it to your data inside your private cloud. The big thing about that is data never leaves." On the AWS side, Jason Bennett, Vice President and Global Head of Startups and Venture Capital, framed it as a permission structure: giving AWS customers the flexibility to run Superblocks in their own virtual private cloud with model choice on Amazon Bedrock means business teams can build AI applications on enterprise data without putting security at risk.

There's one more piece: Superblocks Smart Router, running on Amazon Bedrock, which the companies say cuts inference costs by up to 30%. We'll come back to that number, because there is currently no way to verify it — and there's an interesting reason why.

From Retool's neighborhood to an agent company, in five years

Superblocks was founded in 2021 in New York by Brad Menezes and Ran Ma. Menezes was a lead product manager at Yelp and a senior director of product management at Datadog, and spent time as a Sequoia scout. Ran Ma, now CTO, was an engineer at Morgan Stanley and an engineering lead at Confluent, and before Superblocks co-founded a help-desk SaaS company called Supportive. Back-office finance plus developer tooling — that combination shows up all over the product.

The original idea sounds unglamorous today: internal tools. Large companies burn enormous engineering time building admin panels, order-management screens, and approval workflows that never leave the building, and that nobody wants to maintain. That's the market Retool opened. In August 2022, Superblocks raised $37 million from Kleiner Perkins, Greenoaks, Spark, and Meritech, with angels drawn from the founding and executive ranks of Airtable, Twilio, Okta, Confluent, Firebase, Instacart, Fivetran, Box, Yelp, and DocuSign. Customers named at the time included Motive, Payhawk, Clearco, Papaya Global, and Alchemy.

The pivot came in May 2025. Superblocks raised another $23 million and shipped Clark, an AI agent that builds internal apps from natural language. Clark isn't a single model call — it's a multi-agent setup with specialist agents for Design, IT, Engineering, Security, and QA, and it emits readable React and TypeScript rather than an opaque runtime, applying the company's own design system and access controls automatically. That round brought total funding to $60 million and added names like Box CEO Aaron Levie and Workday founder Aneel Bhusri.

Then things accelerated. Superblocks 2.0 landed on April 15, 2026, already carrying private VPC deployment and Platform MCP — an interface that lets admins programmatically query and act on builders, applications, integrations, permissions, audit logs, and usage events. In that post the company stated flatly that vibe coding had become the number one attack vector inside the enterprise, and named SoFi, Airwallex, and LinkedIn as production users. The August 3 release is that architecture with an AWS stamp on it. Worth noting the scale: roughly 50 employees. At $60 million raised and 50 people, this company isn't winning on size. It's winning on position.

What 3.0 actually sells is what happens after the app is built

Read the 3.0 feature list carefully and something stands out. Very little of it is about generating better apps. Almost all of it is about what happens to an app once it exists: pre-deployment scanning, package-provenance control, vulnerability alerting, policy enforcement. Superblocks has essentially opted out of the generation-quality race and gone to sell the downstream assembly line instead.

The architecture works like this. The platform deploys into the customer's AWS VPC, keeping data, code, applications, and inference inside the existing security perimeter governed by AWS IAM, networking, encryption, and audit policy. When a user asks for an app that needs storage, Superblocks automatically provisions Aurora and S3 resources inside that VPC — and because Aurora supports scale-to-zero, idle apps stop accruing meaningful database cost. All inference runs through Bedrock, restricted to organization-approved models, and the company states that prompts and data are not used for model training.

Security is the headline. A Security Agent Swarm inspects every code change before production deployment, pairing specialized security agents with deterministic scanners. The stated target list is SQL injection, hardcoded secrets, insecure data flows, permission bypasses, and multi-step attack paths. Custom Policy Agents enforce company-specific requirements across every app. On supply chain, the platform connects to private package registries and blocks public NPM access at the network level, maintains a Software Bill of Materials, and continuously scans for CVEs, with a Security Center that alerts app owners and admins when new vulnerabilities are disclosed.

3.0 component AWS service it binds to Problem it claims to solve
Cloud-Prem deployment VPC, IAM, networking, encryption, audit Enterprise data leaving for an outside SaaS
Smart Router Amazon Bedrock Burning frontier-model tokens on trivial tasks
Auto resource provisioning Aurora (scale-to-zero), S3 Business teams waiting weeks for a database
Security Agent Swarm (platform-native) Injection, leaked secrets, permission bypass pre-deploy
Supply chain protection Private registries, SBOM/CVE scanning Malicious NPM packages, unpatched CVEs
Builder MCP MCP clients (Claude, Cursor, Slack, ChatGPT) Editing apps without switching tools

Builder MCP deserves attention too. Developers can iterate on Superblocks apps from inside Claude, Slack, Cursor, or ChatGPT. And the platform opens the door in the other direction as well: you can import work built in Claude, ChatGPT, Lovable, or Replit. That's the go-to-market wedge. When IT belatedly discovers that a business team already shipped something on Lovable, the answer stops being "delete it" and becomes "move it here." The strategy isn't to block shadow IT. It's to absorb it.

Pricing is public, which is unusual and informative. The AWS Marketplace listing shows a 12-month contract at $149,999, with $1.00 per unit for Governed Agent Units and hosted apps beyond the base package, plus private-offer negotiation. The listing carries a 4.7 rating across 191 reviews. That price tag tells you the buyer: this is not an individual developer tool, it's a procurement-cycle product. And buying through AWS Marketplace means it can draw down existing AWS committed spend, so no fresh budget approval is required. Half the practical value of this partnership lives in that purchasing path.

Who gets what out of this

AWS's motive is straightforward. In the frontier-model brand war, AWS trails OpenAI, Anthropic, and Google. So the story AWS keeps telling is that models are interchangeable but the orchestration, security, and data layer above them stays put — on AWS. Superblocks is a clean demo of that thesis. Every app that runs inside a customer VPC spins Aurora, fills S3, and consumes Bedrock tokens. Every additional internal app is an incremental line on the AWS bill. That math justifies a multi-year joint marketing commitment.

For Superblocks this is survival strategy. A 50-person company with $60 million raised cannot go head-to-head with competitors carrying multi-billion-dollar valuations and consumer-scale distribution. But if AWS field sellers open doors at regulated-industry accounts on your behalf, the geometry changes. Even the public "preferred cloud provider" framing does real work: it removes one line item from the risk column in an enterprise vendor review.

The actual audience for this announcement is the customer's CISO and IT organization. Here's the year they've had: a business team built something on Lovable or Replit and started using it, IT doesn't know which database it touches, there are no logs, the person who built it has left, and nobody knows what's in the dependency tree. Superblocks sells a managed corridor instead of a ban. That's not a new argument — it's the exact argument low-code made twenty years ago, which means the familiar failure modes come along with it.

For frontier model providers the news is ambiguous. Menezes's line — customers are done being beholden to expensive frontier models — is aimed squarely at Anthropic and OpenAI. But Smart Router still sends planning and reasoning work to frontier models; what changes is the point of consumption, moving from direct API relationships to Bedrock. If you're sold through Bedrock, the damage is limited. If you sell only through your own API, this hurts more.

How much should you trust "up to 30%"?

Stop here for a second. The Smart Router savings figure is a vendor claim, and there's a coincidence worth flagging. When AWS made Amazon Bedrock Intelligent Prompt Routing generally available on April 22, 2025, the language it used was "reduce costs by up to 30% without compromising on accuracy." Nearly word for word. Public materials don't let you tell whether Superblocks derived its number from the same benchmark or whether "up to 30%" has simply become the industry's default routing headline.

Three things to weigh. First, "up to" is a ceiling, not an average. In AWS's own machine learning blog, testing against open source and internal RAG datasets produced average savings of 63.6% because 87% of prompts routed to a cheaper model while holding baseline accuracy. That cuts both ways: under the right task mix the savings can be far larger, and under the wrong one they can approach zero. The mix decides the outcome.

Second, what exactly is being saved. The 30% applies to inference token cost, not total cost of ownership. The $149,999 annual platform contract, the compute running in your VPC, and Aurora and S3 charges are all untouched. Ask yourself what share of total spend Bedrock tokens actually represent for a portfolio of internal apps, and the absolute dollar figure shrinks fast.

Third, the cost of routing badly. Send a task to a cheap model, get a wrong answer, retry twice, and the savings evaporate — possibly into a net loss. Evaluating that requires routing accuracy, retry rates, and per-task success rates, and none of those have been published. Neither Superblocks nor AWS has released a reproducible benchmark or methodology, and no independent third-party validation exists. Treat the number as a hypothesis, and if you're seriously evaluating, run a two-to-four-week pilot on your own task mix and measure it yourself.

One more unverified figure deserves a flag. Several trade outlets reported that a financial services firm called Flex deployed 70 applications across 18 departments in its own AWS private cloud. That claim appears in neither the AWS press release nor the official Superblocks 3.0 post. The customers those two primary sources actually name are Virgin Voyages and Matthews. So this article won't treat the Flex numbers as established fact.

Twenty years of low-code — who survived and who didn't

The promise that business users will build their own apps is at least three decades old. Lotus Notes in 1989, Visual Basic in 1991, a generation of departmental Microsoft Access databases in the 2000s. Then Salesforce's Force.com in 2007 and the Lightning Platform in 2014, and Microsoft PowerApps in 2015-2016. The cycle repeats: explosive early adoption, then IT discovers thousands of apps nobody can account for, then a governance program starts.

Look at what worked first. PowerApps won because it was welded into Office, Dynamics, and Azure AD. What sold wasn't the ability to build an app — it was that apps automatically inherited existing identity and data. The same logic underwrote Siemens acquiring Mendix for roughly €600 million in 2018 and OutSystems raising at a $9.5 billion valuation in 2021. When Superblocks emphasizes direct wiring into IAM, Bedrock, Aurora, and S3, it is copying that lesson precisely. The moat is the wiring, not the builder.

The failures are just as instructive. Google App Maker was bundled with G Suite and still shut down in 2021. Ecosystem attachment alone isn't enough; without a lifecycle for operating, transferring, and retiring the apps people create, the portfolio becomes a liability. And the whole low-code category kept hitting the same wall. In Gartner research for technical professionals, ineffective governance ranked as the number one issue raised in client discussions about Power Apps and Power Automate, with misuse, solution sprawl, data leakage, and orphaned solutions cited as the headline risks. A KPMG survey of 715 companies in EMEA found 73% of organizations planning low-code adoption had not yet defined governance rules.

There is one way vibe coding genuinely differs from classic low-code. Low-code assembled vendor-defined components, which capped the blast radius. AI agents write real code, pull real packages, and call real APIs. Superblocks putting SBOM tracking, CVE scanning, and network-level NPM blocking at the front of its 3.0 announcement is an admission of that difference. It also matches what the rest of the market is signaling: on August 3, AI agent security startup Zenity closed a $125 million Series C led by Norwest. In the same week, two companies started selling different answers to the same question — who watches what the agents build?

How Lovable, Replit, and Retool counter this

The most direct counter is moving upmarket. Replit raised $400 million in March 2026 at a $9 billion valuation in a round led by Georgian, says it has users at 85% of the Fortune 500 and more than 50 million users total, and told investors it is targeting $1 billion in run-rate revenue by year end. When a company with that funnel decides to bolt on enterprise deployment options and audit tooling, it's a matter of timing, not capability. Lovable is making the same climb from consumer and SMB growth toward enterprise contracts.

The second counter is the hyperscalers' own first-party products, and this may be Superblocks' largest structural risk. AWS already ships Bedrock Intelligent Prompt Routing, which does much of what Smart Router does, as a native service. Microsoft already bundles Power Platform, Copilot Studio, and Azure AI into a single enterprise motion. Google has AppSheet and Vertex. Partnering with AWS neutralizes some of this risk, but a partnership is a contract, not a guarantee — and there is a long history of AWS absorbing well-selling marketplace categories into first-party services.

The third counter comes from the incumbent internal-tools camp. Retool has been building this market since 2017 and still holds a large installed base in exactly the seat Superblocks wants. If Retool hardens its own agent and self-hosted story, Superblocks' differentiation narrows toward "we have a formal AWS relationship." Meanwhile Salesforce and ServiceNow attack from a different angle entirely: don't build an app, just have an agent handle the work inside the platform you already own.

The fourth counter is security specialists. Companies like Zenity occupy a deliberately neutral position — whatever builder you use, we watch it from above. The Security Agent Swarm protects apps built in Superblocks, but real enterprises are a mess of Lovable, Cursor, Copilot, and Claude output all at once. Whether security embedded inside the platform beats security layered on top of every platform is genuinely unresolved.

What actually changes, depending on who you are

Developers should feel two things at once. Fewer tickets asking you to build an admin panel — good. But new work arrives: reviewing code business teams generated, authoring policy-agent rules, curating approved component libraries. That Superblocks emits React and TypeScript matters, because at least you can read and fix it, which is more than opaque low-code offered. Still, "our team now owns 300 AI-written apps" is a new category of debt. The things worth setting up now are an approved package registry and an app-retirement policy.

Enterprise decision makers get three practical takeaways. One, in-VPC execution is a genuinely useful card for clearing procurement in industries with data residency and sovereignty requirements — financial services, healthcare, public sector. Two, buying through AWS Marketplace can draw on existing committed spend, which shortens the budget approval path considerably. Three, the 30% savings figure belongs in a pilot, not a contract. Put it into an ROI model without measuring your own task mix for two to four weeks and you will have an awkward conversation later.

Investors should read the structure rather than the single company. This is a signal that the vibe-coding market is splitting: a growth-rate game in consumer and SMB, and a governance game in the enterprise. The former is already richly valued; the latter is early. But the latter also sits directly in the path of hyperscaler absorption, so discount accordingly. That Superblocks reached this position on $60 million and 50 people cuts both ways — impressive capital efficiency, and plenty of room for a valuation reset at the next round.

Everyday users won't notice anything this week. If you work at a large company, though, you might within six to twelve months. The internal screens you used to file a ticket for and wait months on — inventory lookups, approval queues, departmental reports — start arriving in days. The Virgin Voyages detail in the Superblocks materials is exactly that picture: one senior analyst replacing 15 spreadsheets with a single app, part of a rollout of production apps across seven departments with no dedicated frontend engineers. The open question stays the same, though: what data does that app touch, and who owns it when its author leaves?

🥄 Three Things You're Probably Wondering

— So what does this mean for me? If you're an individual developer or a general user, nothing directly — this is a $149,999-a-year enterprise contract product. But if your job involves building internal tools, there's a good chance the work shifts over the next year or two from "I build it" to "I review what the business team built."

— Does the data really never leave? Architecturally, the platform and the inference both run inside the customer's VPC, calls go out only to organization-approved models through Bedrock, and the company states prompts and data aren't used for training. What public documentation doesn't fully spell out is how much control-plane traffic, license validation, and telemetry crosses the boundary. If you're evaluating this seriously, get that in an architecture diagram and in the contract.

— Is the 30% savings claim real? Too early to call. It's essentially the same phrasing AWS attached to its own Bedrock routing service, and neither company has published a reproducible benchmark. The savings also apply only to inference tokens — platform licensing and infrastructure costs don't move. Until you've piloted it on your own workload, treat it as a reference point rather than a plan.

Further Reading

Numbers are as of announcement and may change.