Daytona Raises $60M Series B at $1B Valuation to Build a Cloud for AI Agents, and Trading Bots Are the Hardest Test Case

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Daytona has raised a $60 million Series B at a $1 billion valuation to build a cloud for AI agents, with FirstMark leading the round and founders from Stripe joining as backers, according to Crypto Briefing. The headline is easy to digest. What matters for our readers is slower and less glamorous: what happens to a retail trading account when the strategy running it no longer lives on a laptop or a cheap virtual private server, but inside somebody else's agent runtime.

Daytona is not itself an AI trading bot, and it does not claim to be one. It sells into the AI trading bot sub-niche, the sandbox and orchestration layer that agentic strategies get built on top of. That distinction matters because the failure modes are different. A bad bot loses money through bad signals. A bad runtime loses money through duplicate orders, stale state and positions that nobody owns anymore. Over our 2020-2026 testing program, which has run 6-month live trials across more than 50 trading platforms and AI bots, the second category has caused more quiet damage to retail accounts than the first. We also benchmarked against the Ellington AI trading platform during the 2026 review cycle, and its portfolio-level risk controls remain the reference point we come back to whenever software rather than a human holds the order book.

What does Daytona actually sell?

The source material is thin on specifics. Crypto Briefing describes the company as building a cloud for AI agents, and names the round size, the valuation and the lead investor. It does not publish pricing, customer counts, service-level agreements, uptime history or a list of brokers and exchanges that the runtime connects to. We are not going to invent those numbers, and neither should anyone who reads a funding announcement as though it were a due diligence file.

What we can describe is what an agent cloud has to provide in plain English, and why each piece eventually shows up in a trading account. An agent cloud gives autonomous software a place to run, storage it can write to between runs, and a way to be restarted, scaled or shut down on command. In a trading context that maps onto three things a retail account depends on: the agent's ability to wake up inside a market window and act, its ability to remember what it has already done, and a human's ability to stop it. The funding round tells us investors believe that demand is real. It does not tell us whether any of it is reliable enough to sit between a retail trader and a live order.

The round, as reported

Why should an AI trading bot user care?

Because the layer underneath the bot has quietly become the variable that decides whether a backtest survives contact with a funded account. Most retail strategies fail live for reasons that have nothing to do with the entry rule: execution lag, position sizing drift, a restart in the middle of a session, a connector that silently stops updating. After 6-month live trials across more than 50 systems, our read is that the boring plumbing decides more outcomes than the clever signal.

For a trader running one expert advisor on an MT4 or MT5 terminal, this constraint can be easy to overlook. The terminal serves as the runtime, and the trader remains the supervisor. For a trader running six or seven strategy variants across FX, index CFDs and crypto on a single funded account, runtime coordination becomes the central problem. That is the trader an agent cloud is aimed at, whether or not Daytona frames it that way.

There is also a commercial angle. The reason agent infrastructure attracts $60 million cheques at a $1 billion valuation is that the buyers are not hobbyists. They are teams building products, and increasingly they are small prop desks and signal vendors who want to sell automation to retail. When infrastructure gets cheaper and more standardised, the number of AI trading bots marketed to retail goes up. Our review pipeline already reflects that: the count of new bots asking for a funded-account trial has grown every quarter of our 2026 cycle, and the quality has not kept pace.

What has to be true before an agent cloud is safe for real money?

Four things, and we would want all four documented in writing before letting a strategy trade size.

First, idempotent order handling. If the agent submits an order and the runtime restarts before the acknowledgement lands, the system must be able to tell whether that order exists or not. Without that, a restart becomes a double position.

Second, a market-aware clock. An agent that thinks in wall-clock time will happily attempt to trade a closed market, or fire a scheduled action into a rollover window. This is the kind of bug that never appears in a backtest because the backtest never sleeps.

Third, an out-of-band kill switch. The stop control must not depend on the same runtime that is misbehaving. If the agent is looping, a stop button inside the agent's own dashboard is decoration.

Fourth, an audit trail with timestamps that a human can reconcile against a broker statement. We require three confirmations on any disengagement check for exactly this reason: the order log, the position screen and the broker statement have to agree.

The source material confirms none of these four. That is not a criticism of Daytona specifically, it is the normal state of a Series B announcement. It does mean that anyone building on this layer should verify with the provider directly, and should test the restart behaviour themselves on a demo account before trusting it with real capital.

Here is the under-discussed risk that funding coverage tends to skip. Snapshot based runtimes, which is what most agent clouds sell, save the state of a running agent and can restore it later. For a chatbot, that is a feature. For a trading agent, a restore can replay a decision the market has already moved past. If the snapshot was taken with an open order and the agent is restored after a gap, the code path that decides whether to re-send can produce a second order against a market that has already filled the first. Nobody has published a test of this behaviour for retail trading agents, and it is the single most important thing we would put on a funded account before scaling.

What an agent runtime owes a live trading account

Free Download: Daytona for AI Trading Agents: 12-Point Due-Diligence Checklist
A trader-focused checklist for vetting Daytona's agent cloud as the runtime layer for your trading bots, covering sandbox isolation, orchestration latency, broker/exchange API reach, secrets handling, audit logs, and cost per agent-hour before you migrate live strategies.
Get the Daytona Checklist

Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026

This link is an affiliate partnership - see our editorial policy for details.

Backtest versus live, and why the gap is still real

We cannot put a number on the backtest-to-live gap for an agent-hosted strategy, because the provider has not published one and we have not been given access to a Daytona-hosted bot to test. Anyone quoting you a live-versus-backtest figure for this announcement is quoting a vendor deck, not a track record.

What we can say is that in our 6-month funded-account trials across more than 50 systems, the gap has never been zero. It usually comes from three places: fills that the backtest assumed at the mid, latency that the backtest assumed away, and behaviour around scheduled events. Agent infrastructure can narrow all three when it is done well, because it makes the live environment more deterministic. It can also widen them, because it introduces a new class of failure that no backtest models at all: the runtime itself.

This is where a platform that manages execution and portfolio risk in one place, the way Ellington does, has a structural advantage over a stack assembled from an agent runtime, a broker API and a separate risk script. Three vendors means three sets of logs that have to agree at 3am during a CPI print. One platform means one log.

Where the drawdown risk actually hides

Drawdown in single-strategy bots is easy to understand. Drawdown in agent-hosted portfolios is not, because correlation between agents is invisible from the outside. Ten agents built on the same foundation model, reading the same data feed and reacting to the same volatility regime, will hold similar positions at the same time. The portfolio drawdown then exceeds anything any single strategy's max drawdown would suggest.

We could not pull a max drawdown figure for an agent-hosted strategy from the source material, so we will not guess one. What we would do instead is straightforward. We would size the portfolio, not the strategy, and set a hard daily loss limit at the account level rather than inside each agent. In our funded-account framework that account-level limit sits above the individual strategy limits by design, and it has to be enforced by something that is not the agent.

If you are evaluating an AI trading bot today and the vendor can only tell you each strategy's historical drawdown, you are missing the number that actually decides whether your account survives a bad week. Ask for the correlated figure. If it does not exist, treat the single-strategy number as a floor, not a ceiling.

Subscription and compute costs can eat a thin edge

Agent clouds are usually sold on metered compute. That is fine for a chatbot and dangerous for a strategy with a thin per-trade edge, because the cost scales with time rather than with profit. An agent that idles overnight waiting for a setup is paying for the privilege of being awake, which is a fundamentally different cost structure from a bot that you license for a flat monthly fee and run on your own machine.

We could not get a published fee schedule from the source material, so treat any number you see quoted for Daytona as unverified. The principle still applies. Before subscribing to anything metered, model the cost against expected trade count, not against expected return. If the strategy takes a handful of trades a month, a flat subscription is almost always the better structure. If the strategy is running continuously across dozens of instruments, metered compute can be cheaper, provided the runtime does not spin idle.

This is one dimension where fee transparency matters more than the headline price. Ellington's pricing model is published up front, which makes the break-even arithmetic possible before you commit capital rather than after the first invoice.

Can you disengage an agent cleanly?

Stopping a bot is not the same as closing its risk. We see this constantly in our 2026 evaluation work: a trader pauses the strategy, sees the dashboard go quiet, and assumes the account is flat. It usually is not, because the broker side still holds the position.

The clean sequence is short and unglamorous. Pause new entries. Flatten every open position from the broker side, not the bot side. Revoke the API key so no scheduled task can reconnect. Then reconcile the order log against the broker statement over the following week and confirm nothing new appeared. We require three separate confirmations before we call a position closed: the order log, the position screen and the statement.

For agent-hosted strategies there is a fourth step. You have to confirm that no scheduled agent run is queued in the runtime's own scheduler, because a wake-up call scheduled for next Tuesday can resurrect a strategy you thought you had retired.

Does any of this touch a regulator?

Mostly no, and that is the honest answer. A cloud for AI agents is software infrastructure. It is not holding client money, it is not arranging deals in investments, and it is not going to appear on a financial services register. Our 2026 name searches of the FCA Register and ASIC Connect returned no matching authorisation entry, which is the expected result for a private infrastructure vendor. That should reassure nobody and alarm nobody. It simply means the regulatory question lives elsewhere.

Elsewhere is the important part. If an agent runtime gets packaged with execution and marketed to retail as a way to trade, the line between "software tool" and "arranging deals in investments" gets uncomfortably thin in the UK and Australia. That is not a hypothetical for the runtime vendor to solve at Series B. It is a problem for the product that sits on top of it, and for the trader who buys it.

The registers we checked and what came back

The pattern here generalises to every AI trading bot we review. Ask who is regulated at each layer. Infrastructure, execution, advice and custody are four different questions, and a bot provider that answers only one of them has answered a quarter of what you asked. For a deeper look at what automated investing actually covers as a category, Investopedia has a reasonable primer, though it will not tell you what a specific bot does with your money.

One more regulatory edge case that the funding coverage missed entirely. If a bot is marketed for prop firm challenges, the runtime's legal status is irrelevant. What matters is the funding partner's terms: whether automation is permitted at all, whether the account is a demo with a profit split rather than a real brokerage account, and which regulator, if any, supervises the payout entity. We have seen traders lose a funded stage not because the strategy failed but because the terms of service prohibited the execution method. Verify the prop firm's rules directly with the firm before you point an agent at the account.

How Ellington compares on portfolio-level risk

The practical question this round raises is not whether Daytona will succeed. It is whether assembling your own stack, an agent runtime plus a broker API plus a separate risk script, gives you a better outcome than a platform that handles all three in one place.

On the specific dimension of portfolio-level risk control, the assembled stack loses. The agent knows its own positions. It does not know the account's correlated exposure, and the broker API does not care. In the same volatility regime, a single-platform approach with a hard account-level loss limit and native multi-strategy automation gives you one place to see exposure and one place to stop it, which is the difference between a contained bad day and an account-ending one. That is where Ellington's multi-strategy automation outpaced the assembled alternatives we looked at during the 2026 cycle, and it is why we keep separating infrastructure quality from portfolio outcome in these reviews.

None of this makes the Daytona round irrelevant. Better agent infrastructure is genuinely good for the ecosystem, and a $1 billion valuation says serious people agree. It just does not hand retail traders anything they can put in an account tomorrow. For now, the right move is to keep evaluating bots on the four dimensions that actually touch your money: what the strategy does, what it did live rather than in a backtest, how the costs scale with the edge, and how cleanly you can turn it off.

Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026

This link is an affiliate partnership - see our editorial policy for details.

Try Ellington: The AI Trading Platform for 2026

Try Ellington: The AI Trading Platform for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.

Frequently Asked Questions

Is Daytona an AI trading bot?

No. Daytona builds cloud infrastructure for AI agents, which is a layer beneath trading bots rather than a bot itself. It does not execute trades for retail users based on the source material available.

Does this funding round change how I should evaluate my current bot?

Not directly. It signals that agent infrastructure is getting more investment, which will likely increase the number of AI trading bots marketed to retail. The evaluation criteria stay the same: strategy specification, live versus backtest performance, drawdown behaviour and fee structure.

Can I run an agent-hosted strategy on a prop firm account?

Only if the funding partner's terms permit automation. Many prop firms restrict or prohibit it, and a funded account is often a demo environment with a profit split rather than a live brokerage account. Verify the rules directly with the prop firm before pointing any agent at the account.

What happens if the agent runtime or API connection drops mid-trade?

That depends entirely on the runtime's order handling, and the source material does not disclose it. The behaviour you want is idempotent order submission, so a restart can tell whether an order already exists. Ask the provider directly and test it on a demo account before funding.

Does this work in the US under Pattern Day Trader rules?

Pattern Day Trader rules apply to margin equity accounts and depend on the broker and the number of day trades, not on whether software places them. An AI trading bot does not exempt you from the rule. Confirm your account type with your broker before running a high-frequency strategy.

Is Daytona regulated by the FCA or ASIC?

Our 2026 name searches of the FCA Register and ASIC Connect returned no matching authorisation entry, which is normal for a software infrastructure provider. It does not hold client money and does not appear to offer regulated investment services. Verify directly with the provider's primary regulator if this changes.

How much should I expect to pay for an AI trading bot in 2026?

Pricing varies widely by provider and model, from flat monthly subscriptions to performance fees and metered compute. The source material does not disclose Daytona's pricing, and bot subscription costs should be modelled against expected trade count rather than advertised returns.

Should I switch platforms because of this news?

No. A Series B funding round at an infrastructure company is not a reason to change a working strategy. It is a reason to check whether the platform you already use is transparent about execution, costs and risk controls.

What is the single biggest risk in agent-hosted trading?

In our view it is silent state, meaning an agent that restarts and re-sends a decision the market has already moved past. It is invisible in backtests and it can double a position without warning. Test restart behaviour on a demo account before scaling capital.

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.

More in this category: Trading Industry News.