Data center growth is now limited by power, land, water and public consent rather than by demand. Most processors spend much of their time idle, waiting. F1R3FLY concurrency gives them useful work during those waits, so the same hardware, on the same power budget, does considerably more.
The basic structure of modern computing was set out by John von Neumann in 1945, and it keeps memory and processing apart. Hardware has changed beyond recognition since, but that separation remains, and data centers reproduce it at the scale of a building: processing sits in one area, storage in another, and data travels between them constantly.
Around twenty years ago, physical limits leveled processor clock speeds off at roughly 3–5 GHz. Since then, waiting on data has become the main limit on most data-intensive work. A processor sees a program as a list of instructions to be worked through one at a time, like a recipe followed by a single cook, and most of the time that cook is standing idle: waiting for data to arrive from memory, for a lock to be released, or for another task to finish.
The industry's response to rising demand has been to build more, so today's enormous facilities deliver their computing power through sheer scale rather than through efficiency.
Because the Rho-Calculus makes dependencies between processes explicit, independent work can proceed while other work waits. The processor is given something useful to do during the idle moments, so more of the hardware is working more of the time. Energy moves in the same direction: an idle chip still draws a large share of its full power, so keeping it busy raises the useful work done for each unit of energy, even though total power draw rises slightly.
How large the gain can be is set by how much time is wasted today. A processor busy half the time can at most double its output; one busy a tenth of the time can at most raise it tenfold. The gains are largest where coordination dominates – many-step AI agents, orchestration across models and tools, data pipelines, and any workload made of many small interdependent tasks. They are smallest in the dense arithmetic at the heart of AI models, which is already highly parallel and already keeps specialized chips busy. Concurrency does not make that arithmetic faster. It makes sure the hardware around it is not left waiting.
There is one real condition. A processor can only run a program concurrently if the program is written to allow it. Software written in conventional languages is sequential at heart and gains little, falling back to the same one-step-at-a-time processing it would receive anywhere else. The gains described here depend on workloads being written for concurrency, ideally in F1R3LANG, which expresses a program as communicating processes from the outset. That rewriting has a real cost, paid once. The returns recur every year.
Each builds on the last. All assume the workloads concerned have been written for concurrency, and all use the same reference point: a typical 40 MW AI facility that sells computing capacity, costing around $1.5 billion to build and around $340 million a year to own and run, earning around $110 million in profit.
| Scenario | Throughput | Revenue | Costs | Annual profit |
|---|---|---|---|---|
| Today | 1× | $450M | $340M | $110M |
| 1. Software only | 1.5× | ~$675M | ~$350M | ~$325M |
| 2. Memory pooled beside processors | 3× | ~$1.1B | ~$400M | ~$700M |
| 3. New memory-centric chips | 20× | ~$3.6B | ~$0.9B | ~$2.7B |
Throughput figures are central estimates for workloads rewritten for concurrency, not upper limits. Revenue rises by less than throughput because the price of compute is assumed to fall as capacity grows – by around a fifth in Scenario 2 and around 60% in Scenario 3 – with most of the gain passing to customers. Cost basis: Epoch AI's model of a one-gigawatt AI data center, scaled to 40 MW. These are scenario-dependent estimates, not forecasts or guarantees.
The building and its hardware stay exactly as they are. Rewritten workloads recover the idle time, lifting throughput by around half. Because the servers are already paid for, almost all the extra output becomes profit. Existing facilities could absorb around half as much work again before new ones are needed.
Pooled memory is added at rack level using current standards, so large working sets sit close to the processors rather than across the network. Rewiring and around $200 million of equipment, far less than a rebuild, for around three times today's throughput on suitable workloads.
The floor moves to chips that place large amounts of memory directly beside the processing cores, which matches how F1R3FLY concurrency works: each channel and its processes sit on a core and its local memory, so data rarely moves. Around ten times from the hardware, two to five times from concurrency in the work around the models, giving a range of roughly 10× to 50× with 20× as the central case.
Concurrency will arrive workload by workload, as programs are written for it. The obvious place to begin is artificial intelligence: AI-optimized servers are forecast to account for 31% of global data center electricity in 2026, to use more than conventional servers by 2027, and to reach around half of all data center work by 2030.
Compute is among an AI company's largest expenses and a main limit on how fast it can grow, and most of that cost is the hardware itself rather than the electricity to run it. Running 4,000 MW of AI capacity costs around $34 billion a year. The table shows what each scenario would save on the cost of doing the same amount of work.
| Improvement in work per dollar | Annual saving per 4,000 MW |
|---|---|
| 1.5× (software only) | ~$11B |
| 2× (reference case) | ~$16B |
| 3× (pooled memory) | ~$21B |
| 20× (new chips) | ~$29B |
A saving of around $16 billion for every 4,000 MW therefore needs only about a twofold improvement in the work done for each dollar. That sits between the first two scenarios, far below the central case for new chips. Savings are stated before the one-off cost of rewriting software, which each company would weigh against them.
Data centers worldwide use around 65,000 megawatts of electricity on average, of which AI accounts for roughly 20,000. Where a facility does around twenty times the work on the same power, the energy needed for each unit of work falls by around 95%. If today's AI work were all moved to converted facilities it would need around 1,000 megawatts instead of 20,000, with a range from about 4,000 megawatts if the gain is only fivefold to about 200 at the most optimistic.
Not all work will be rewritten, and the more important figure is what happens to growth. A doubling of computing demand could be met with around 3,000 additional megawatts rather than another 65,000. Water use for cooling falls broadly in line with energy, and further still where waterless cooling is used, and fewer new facilities means less land taken.
This is the part that reaches beyond any single operator's balance sheet. Computing power could keep growing without the energy, water and land costs that now make each new data center a matter of public argument – and the operators who move first capture the advantage while everyone else is still buying megawatts.
The estimates here rest on three measurable claims, and each can be tested on a named workload rewritten in F1R3LANG before anyone commits to anything larger:
A pilot that measures those three figures shows which scenario applies, and to which workloads. The economics hold even if the results come in well below the central estimates: at only five times today's throughput, with prices falling by around 30%, the reference facility would still earn around $630 million a year, nearly six times today's profit.
No scientific breakthrough is required. The hardware exists and the model is proven in production.
If you operate data center capacity, or buy a great deal of it, a measured pilot on one workload is the place to start.
Get in Touch