Your infrastructure is part of your strategy
Most desks treat infrastructure as an IT line item. In continuous intraday trading it is part of the strategy. The exchange processes orders one at a time, in the order they arrive. How fast your order arrives, and whether your stack is up at all, decides whether the strategy you designed is the strategy you actually trade.
Slow orders pay the fast ones
The best measurement of what latency costs comes from the UK Financial Conduct Authority. Economists at the FCA and the University of Chicago analyzed the full message data of the London Stock Exchange, including the orders that failed. That let them see not only who won each race to the order book, but who lost.
The results, published in the Quarterly Journal of Economics:
Races for the same price happened about once a minute per stock and were usually decided within 5 to 10 microseconds.
About 20% of all trading volume took place in these races.
Six firms accounted for over 80% of race wins and losses.
The slower side paid a tax of roughly 0.42 basis points on all volume traded, about $5 billion a year across global equity markets.
This was equities, but the mechanism is identical to EPEX SPOT continuous intraday: a central order book with price-time priority, processed serially. When a plant trips, every algo sees the same signal. The first order to reach M7 takes the liquidity. The rest get the next price level. EPEX SPOT reports that around 84% of intraday trades now arrive via API, and since October 2025 each day splits into 96 quarter-hour products. More races, run by machines, over thinner books.
Shared hardware adds variance you cannot see
Latency is not one number. On shared or undersized hardware it is a range, and the tail of that range is where you lose fills.
We measured this on Voltnir itself, 12 hours each against the EPEX SPOT simulation environment:
Order-book delta applied: 12 μs on a dedicated 8-core machine, 30 μs on a shared 2 vCPU VM.
Order in to M7 wire: 118 μs dedicated, 480 μs shared.
Same binary, same code. The wait for the publisher went up because on a shared VM the process competes with other workloads for CPU time. These are averages. The spikes are worse, and they come exactly when the machine is busiest: when the book is moving.
A cloud-hosted SaaS gateway adds a public-internet round trip to a vendor region on top of that, typically several milliseconds and often tens. In a market where races are decided in microseconds, that is not a small overhead. It is a decision to lose the race.
The exchange was up. The vendor was not.
On 19 July 2024, a faulty update from security vendor CrowdStrike crashed Windows machines worldwide. EPEX SPOT reported normal market status for most markets that day, and order books closed on time. The exchanges were open.
Many energy traders still could not trade. Trayport, the screen most European wholesale power and gas trading runs through, reported infrastructure issues caused by a global outage at a third-party service provider. EEX told clients that those trading via Trayport were having problems as a result. Traders at Shell, BP and Vitol told Reuters their operations were affected. One trader told ICIS they could trade but not book deals.
None of those desks chose to run CrowdStrike through Trayport. They chose a vendor, and inherited that vendor's vendors.
The cloud has the same problem one layer down:
20 October 2025. A DNS failure in AWS's US-EAST-1 region took Coinbase and Robinhood down for hours. Users could not log in, place orders or withdraw. The markets they route to were trading normally.
November 2025. A three-hour Cloudflare outage took down multiple forex and CFD brokers. Finance Magnates estimated the average broker lost about $1.58 billion in trading volume, close to 1% of monthly trading revenue.
8 May 2026. A thermal event in a single AWS data center took Coinbase down for more than two hours. Coinbase had designed its systems to survive the loss of one availability zone. It did not.
Coinbase runs its own engineering organization and still could not design around its cloud provider. A desk renting a SaaS gateway has less control than that: it cannot see the vendor's architecture, cannot choose their region, and finds out about the outage from its own traders.
What this means for risk
In continuous intraday power, an outage or a slow fill is not a missed opportunity you can shrug off. The position is still there. If you cannot close it before gate closure, it settles against the imbalance price, which the TSO sets and you do not.
That makes the infrastructure question a risk question:
Who controls the path to the exchange? Every vendor and region between your algo and M7 is a dependency you cannot monitor or fix.
Who decides when it changes? CrowdStrike was an update nobody at the trading desk approved. Software on your own hardware changes when you decide, after you have tested it.
What happens when something breaks? You want the system to refuse new orders against a stale book, and to always let you cancel.
This is why infrastructure is part of your strategy,
More on the design: why Voltnir runs on your infrastructure.