Durable execution has gone from a niche backend idea to a funding category. On 30 September 2026, Berlin-based Restate announced a $20 million Series A led by Singular, with Redpoint Ventures and Capital One Ventures joining. Two weeks earlier, its much larger rival Temporal raised $550 million at a $12.55 billion valuation. Both companies sell the same basic promise: if a long, multi-step process crashes halfway through, it picks up exactly where it stopped instead of starting again or leaving a mess.
That promise matters more now because AI agents run long, unpredictable chains of model calls, tool calls, waits and approvals. “It never was built for agents in the beginning, but it just happened to be a perfect match for all these problems that agents surface,” co-founder Stephan Ewen told TechCrunch.
This article explains what Restate announced, what durable execution does, why Restate says the category is priced too high, how its design differs, and what a business building agents should check. Our arithmetic is labelled as ours throughout. For wider context, see our guides to execution governance for AI agents and to where agents run in managed agent environments.
Table of contents
- What Restate Announced on 30 September
- What Durable Execution Actually Does
- Restate’s Case: Durable Execution Is Priced Too High
- How Restate Builds Durable Execution Differently
- Restate Versus Temporal in Durable Execution
- Who Uses Restate’s Durable Execution
- What Durable Execution Means for Businesses Building Agents
- What We Still Don’t Know About Restate
- Durable Execution FAQ
- References and Further Reading
What Restate Announced on 30 September
The news arrived in two places at once. TechCrunch’s Marina Temkin published the story at 14:27 UTC, and Ewen posted a longer announcement on the company blog titled “Restate raises $20M Series A to make Durable Execution a building block for every backend”.
The round and the investors
Singular, a European venture firm, led the round. Redpoint Ventures, which led Restate’s $7 million seed round in June 2024, returned. Capital One Ventures, the venture arm of the US bank, also took part. Neither source gives a valuation. Adding the two disclosed rounds gives $27 million raised in total, which is our sum rather than a figure the company published.
What the money is for
According to TechCrunch, Restate will hire a go-to-market team, add engineers and expand its office in the San Francisco Bay Area “to be closer to many of the current and potential customers”. Ewen’s post frames the goal more broadly: to make durable execution “a widespread building block for backends, rather than something confined to expensive, high-overhead workflow runtimes”.
The traction behind it
Ewen told TechCrunch that Restate has closed “multiple six- and seven-figure customer contracts” in recent months. Named users include the coding platform Replit, the start-up DOSS and an unnamed Fortune 500 bank.
| Item | Detail | Source |
|---|---|---|
| Round | $20 million Series A, 30 September 2026 | Restate blog, TechCrunch |
| Lead | Singular | Restate blog |
| Also investing | Redpoint Ventures, Capital One Ventures | Restate blog |
| Earlier funding | $7 million seed led by Redpoint, June 2024 | Redpoint |
| Total raised | $27 million | Our sum |
| Valuation | Not disclosed | n/a |
| Founders | Stephan Ewen, Igal Shilman, Till Rohrmann | TechCrunch |
| Named users | Replit, DOSS, a Fortune 500 bank | Restate blog |
What Durable Execution Actually Does
Most business software runs in steps: charge the card, reserve the stock, send the email. If the server dies between step two and step three, someone has to work out what already happened. Durable execution moves that job from the developer into the runtime.
A crash in the middle of step four
Picture a refund process with six steps. The server restarts during step four. Without a durable runtime, the process either starts again from step one, risking a second refund, or stops and waits for a human to untangle it. With durable execution, the runtime has recorded that steps one to three finished and what they returned, so it resumes at step four with the same inputs.
Journals, replays and exactly-once effects
The mechanism is a journal. Each completed step is written to a durable log before the program moves on. After a failure, the program replays its journal instead of repeating the real work, so a payment call that already succeeded is not called twice. Ewen describes the result as a runtime that “takes responsibility for recording progress, recovering from failures, retrying work, and recovering partial progress”.
Why agents made durable execution urgent
A single agent task can involve “hundreds of model calls, tool calls, waits, retries, callbacks”, Ewen writes, and may need a human to approve a step. Model calls are slow and cost money, and tool calls can move money or delete data. Repeating them after a crash is expensive at best and harmful at worst. Our report on the Claude Code incident that deleted 48,000 files shows how quickly an agent’s actions become irreversible, which is exactly the class of problem a replayable history helps you understand afterwards.
Restate's Case: Durable Execution Is Priced Too High
The most interesting part of Ewen’s post is not the funding. It is an argument about price, and it explains why Restate thinks the durable execution category is still too small.
Tens of dollars against about a dollar
“Managed Durable Execution is commonly priced in the tens of dollars per million durable actions (often $25 – $50),” Ewen writes. “In comparison, databases and queues commonly cost around ~$1 for million writes or events.” He accepts that a durable action does more than a database write, but argues that “at high per-action prices, developers naturally use durability sparingly and create coarse activity boundaries”.
Checking the $25 to $50 range
The range holds up against published list prices. Temporal Cloud’s pay-as-you-go page charges $50 per million actions for the first 5 million each month, falling in steps to $25 per million between 100 and 200 million. AWS Step Functions Standard Workflows charge $0.000025 per state transition in US East, which is $25 per million. Step Functions Express Workflows, a lighter service without the same long-running guarantees, charge $1.00 per million requests plus duration.
Published list price per million units, US dollars (sources: Temporal pricing page, AWS Step Functions pricing, Restate blog)
What rationing looks like in a bill
To show why the gap changes design, we priced a hypothetical agent product on Temporal’s published tiers. It runs 100,000 agent tasks a month. A coarse design records 150 durable actions per task, or 15 million a month. A fine-grained design, where every inference and tool call is durable, records ten times as many: 150 million a month. Only action charges are counted; storage, support and any committed-use discount are left out.
The coarse design costs $675 a month in actions ($250 for the first 5 million, $225 for the next 5 million, $200 for 5 million more). The fine-grained design costs $4,675. At Ewen’s database-like rate of about $1 per million, the same two designs would cost $15 and $150. That thirty-fold difference at scale is the whole of Restate’s pitch in one number.
Illustrative monthly action bill for 100,000 agent tasks (our arithmetic from published tiers)
Latency is the second cost
Price is only half of it. “Getting durable actions from 25+ms to less than 5ms makes the decision to introduce an extra step easier,” Ewen writes. Our arithmetic again: if 1,500 durable steps in one agent task run one after another, 25 milliseconds each adds 37.5 seconds of waiting. At 5 milliseconds, the same steps add 7.5 seconds. For a coarse task with 150 steps, the overhead falls from 3.75 seconds to 0.75 seconds.
Added waiting per agent task if durable steps run in sequence, seconds (our arithmetic)
Inside the loop or around it
The practical difference is where you put durable execution. Around the loop, you know an agent task ran and whether it finished. Inside the loop, you have a record of every inference, tool call and guardrail check. Ewen argues the second gives “a consistent execution history”, which matters “for agents performing sensitive, long-running, or expensive tasks”. Cheap, fast durable execution is what makes the inside-the-loop design affordable.
How Restate Builds Durable Execution Differently
Ewen’s post lists three conventions the team chose to break. They are design claims by the vendor, not independently tested results, but they explain the price argument.
Rule one: build your own storage
Restate “half-broke” the rule against writing your own storage. It uses RocksDB on each node and object stores for snapshots, but built its own replicated log, consensus and failover, and query engine around them. The hot path, making a step durable, “boil[s] down to a quorum write (a single network roundtrip)”. Fewer moving parts is also the operational pitch for self-hosted durable execution.
Rule two: push instead of poll
Most workflow engines put work on task queues and let workers pull it. Restate pushes invocations to durable functions running in containers or on serverless platforms such as AWS Lambda, Cloudflare, Vercel and Cloud Run. Ewen admits this is harder, because pull-based workers get back-pressure for free, so the dispatcher needed its own flow control.
Rule three: state outside the workflow
Restate’s Virtual Objects are “a stateful entity keyed by an ID, for example a chat session, with durable state and one write handler running at a time per key”. Ewen says they have become “a fan-favorite” for agent context, session stores and custom queues. For agents, that means the conversation memory and the durable execution of its steps sit in the same system.
| Design choice | Common workflow engine, as Ewen describes it | Restate’s approach |
|---|---|---|
| Storage | External database plus queue and metadata store | Own replicated log over RocksDB and object storage |
| Dispatch | Workers poll task queues | Runtime pushes invocations to functions |
| State | Kept in a separate database | Virtual Objects keyed by ID |
| Deployment | Several services plus a database | Single binary that forms clusters; self-host or BYOC |
| Target latency per step | 25ms or more | Under 5ms |
Restate Versus Temporal in Durable Execution
TechCrunch framed the round as Restate taking on Temporal, “the heavyweight in the durable infrastructure space”. The gap in size is large, and the numbers are worth setting side by side.
Temporal’s two rounds this year
Temporal announced a $300 million Series D at a $5 billion valuation on 17 February 2026, then a $550 million Series E at $12.55 billion on 14 September, co-led by Lightspeed with Wellington Management, Goldman Sachs Alternatives and Tiger Global. That is a 2.5-fold rise in valuation in seven months, by our arithmetic. Temporal says it processed 1.9 trillion billable actions in August alone and has more than 4,300 paying customers, including OpenAI, Snap and NVIDIA.
Where Temporal came from
Temporal’s founders, Maxim Fateev and Samar Abbas, built workflow systems at Amazon (Simple Workflow Service), Microsoft (the Durable Task Framework) and Uber (Cadence) before starting the company in 2019. Restate’s founders come from stream processing instead: Ewen co-created Apache Flink and was chief technology officer of Data Artisans, which Alibaba bought in 2019 and renamed Ververica.
| Measure | Temporal | Restate |
|---|---|---|
| Founded | 2019 | 2022 |
| Latest round | $550M Series E, 14 Sep 2026 | $20M Series A, 30 Sep 2026 |
| Valuation | $12.55 billion | Not disclosed |
| Scale disclosed | 1.9 trillion billable actions in August; 4,300+ paying customers | Six- and seven-figure contracts; customer count not given |
| Headcount | 570 | Not disclosed |
| Server licence | MIT | Business Source License 1.1, converting to Apache 2.0 after four years |
Size of each company’s latest round, US$ millions
A note on the licence
Restate’s server code is public on GitHub but is not open source in the strict sense. It uses the Business Source License 1.1, which lets anyone run it except as a public “Restate platform service” for third parties. Each version converts to the Apache 2.0 licence four years after release. For most companies self-hosting inside their own network, that restriction will not bite, but it is worth a line in any procurement review.
Who Uses Restate's Durable Execution
Three customer stories in Ewen’s post show the range of jobs the product does. They are the vendor’s accounts, not case studies we have checked.
Replit moved its agent onto it
Replit “moved the execution of the Replit Agent onto Restate as part of a larger evolution of the Agent architecture”. The new design “uses more than 10x more durable actions as the previous generation”. That is the inside-the-loop pattern described above: individual tool calls and inference calls become durable without a round trip to a separate workflow orchestrator.
DOSS replaced a queue
DOSS “runs high volumes of enterprise workflows and subworkflows on Restate”, replacing an architecture built on BullMQ, a popular Node.js job queue. The claimed gains were latency, failure handling, observability and easier debugging.
A bank needed multi-region consistency
An unnamed Fortune 500 bank adopted Restate “for financial workflows with multi-region consistency requirements”. Capital One Ventures joining the round is not proof that the bank is Capital One, and neither source says so.
Compliance for regulated work
Restate Cloud reached SOC 2 Type I in February 2026, with a Type II audit then under way, and offers business associate agreements for HIPAA work to enterprise customers. Ewen names loan origination and medical bill disputes as agent jobs where data must stay “within a network boundary”, which is why self-hosting and bring-your-own-cloud options matter.
What Durable Execution Means for Businesses Building Agents
Most companies will meet durable execution through a platform rather than buy it directly. Still, the choice shapes cost, reliability and audit trails, so it helps to know when it is worth the effort. It sits alongside wider cloud computing choices such as where workloads run and who holds the data, and next to workflow automation tools that handle simpler, shorter processes.
When you need durable execution
You need durable execution when a process is long, touches money or customer records, calls outside services that can fail, or waits for a human. An agent that books, pays, files or deletes belongs here. So does anything a regulator may ask you to explain step by step.
When you probably do not
A short, stateless request that can simply be retried from the start does not need it. Neither does a chatbot that only answers questions. Adding durable execution there adds cost and another system to run for little benefit.
| Question for a vendor | Why it matters |
|---|---|
| What is the price per million actions at our expected volume? | Decides whether you can make every agent step durable |
| What is the added latency per durable step? | Multiplies across hundreds of steps per task |
| Can we self-host or run it in our own cloud? | Keeps regulated data inside your network |
| What licence covers the server? | Affects resale and managed-service use |
| Can we replay and inspect a failed run? | Turns an incident into an audit trail |
| Which audits and agreements are in place? | SOC 2 reports and HIPAA agreements shorten security review |
Measure before you migrate
If you already run a workflow engine, count the actions a typical agent task records today and how long each takes. Those two numbers let you test a vendor’s pricing and latency claims against your own traffic instead of a slide.
What We Still Don't Know About Restate
The round answers some questions about Restate’s durable execution business and leaves others open.
Its own price and speed
Ewen’s post criticises the market’s pricing but does not publish Restate’s rate per million durable actions, and the sub-5-millisecond figure is described as a target rather than a benchmark result. Buyers should ask for both in writing.
How big it is
There is no valuation, revenue figure or customer count. “Six- and seven-figure contracts” describes deal size, not how many deals there are.
A spelling slip in the coverage
TechCrunch’s story calls the co-founder “Stephen Ewen” in its first line and “Ewan” twice further down. The company’s own post and his Apache Flink history use Stephan Ewen, which is the spelling we use here.
Durable Execution FAQ
What is durable execution?
A way of running multi-step code so that each completed step is recorded. If the process crashes, it resumes from the last recorded step instead of starting again or repeating actions such as payments.
How much did Restate raise?
$20 million in a Series A announced on 30 September 2026, led by Singular with Redpoint Ventures and Capital One Ventures. It follows a $7 million seed round in June 2024.
How much does durable execution cost?
Published prices vary. Temporal Cloud charges $50 per million actions for the first 5 million each month, falling to $25 at higher volumes, and AWS Step Functions Standard charges $25 per million state transitions. Restate has not published a comparable per-action price.
Is Restate open source?
Its server code is public but uses the Business Source License 1.1, which bars offering it as a public managed service. Each version becomes Apache 2.0 four years after release.
Who founded Restate?
Stephan Ewen, Igal Shilman and Till Rohrmann, who worked together at Data Artisans and Ververica on Apache Flink.
References and Further Reading
Restate: Restate raises $20M Series A to make Durable Execution a building block for every backend
TechCrunch: Restate lands $20M as the need for durable infrastructure increases with AI agents
Temporal: Temporal raises $550M at a $12.55B valuation
Restate Cloud: SOC 2, Enterprise SSO, and HIPAA
Restate server licence on GitHub
Redpoint: Restate raises $7M to build a next generation durable execution platform
