● ENGINE LIVESign inCreate an account
We found a way to make silicon do more with the work it is given. It runs in software today, on hardware you already own.
We are building the processor it was designed for.
“WUBBERY wins, and we’ll never stop growing until the wheel we just reinvented is attached to the vehicle that is driving innovation for the entire world.”Tim Harkin, founder
Most infrastructure bolts on a component. WUBBERY adds a representation — the effort each request actually carries — and everything downstream reads it: routing, memory, prefetch, governance, anomaly detection, forecasting, even generated media. Each is somebody’s whole company; none of them shared a substrate until now. That is the breakthrough — and it makes the AI you already run faster, sharper and cheaper in any industry that touches it. A measured 32.9% cost cut from routing alone, and a 3.67 µs decision are two of the numbers it produces. Live, published-benchmark, estimated, projected and sample figures are labelled separately throughout this page.
Get WUBBERY or get left behind.
PUBLISHED BENCHMARK 2026-07-30 · i5-1135G7 @ 2.42GHz · 1 THREAD · BEST-OF-5 × n=100,000
REPRODUCE: GET /v1/bench/overhead
Hundreds live today. You run the ones that pay.
Nobody switches all of them on, and nobody should. Tell our team where you want to improve and WUBBERY will get you there. For every 100 you spend on compute today, the ones that pay for themselves are worth 206.4 — and you keep 154.8 of it. Everyone else sells you one of these.
Yes, that is more than 100% of your bill. Here is why it is allowed.
Saving money and making the same hardware do more are two different quantities. Add three cost cuts against one bill and you get nonsense — 123.94% on paper, 86.58% when actually measured, and we publish the smaller one. But money you no longer spend plus work you can now do without buying anything are separate. Added once, they come to 206% of what the fleet costs you today.
We used the conservative figure. The median gain is 2.20×. The mean is 8.49× and the best single module is 90×. Using the mean would roughly quadruple the number on this page.
The whole calculation, so you can check it
- Your compute bill
- Whatever it is — call it 100
- 100
- Cost avoided
- 86.58% measured, three levers composed
- 86.6
- Capacity gained
- 54.50000000000001% less compute for the same work ⇒ 2.20× per unit of compute
- 119.8
- Total value
- 206% of what you spend today
- 206.4
- WUBBERY 25%
- Find nothing, pay nothing
- 51.6
- You keep
- 75% of the value created
- 154.8
- What you still spend
- 13.4 compute + 51.6 our share
- 65
- Cash saved, after paying us
- and 2.20× per unit of compute, which is not cash
- 35
What this is not. The dollars are modelled on a fleet size we chose and a public rate — not measured on your bill. Your traffic, your number, which is why the first thing we do is run it on yours.
Every one of them, measured.
Not a roadmap and not a category list. Each module is run against a named alternative at a fixed seed, and either it stops a failure or it beats the baseline by a measured margin.
Everyone else asks a model. We already know.
Before any work starts, something has to decide what to do. That decision is pure overhead and you pay it on every request. Lower is better; the scale is logarithmic, because at a linear scale ours has no width at all.
WUBBERY measured 2026-07-30 on an i5-1135G7 at 2.42 GHz, one thread, best-of-5 over n=100,000 — a 15 W laptop part, so your rack will beat it. Every other row is that approach’s own published overhead, named by category rather than by company. Independent; no endorsement implied.
Found a new problem? Tell us. That is the whole product.
Hundreds are live today, and that is this month’s number rather than the ceiling. The reason we can add yours in days is that the hard part — the substrate every one of them already runs on — is built, measured and in production. Tell our team where you want to improve and WUBBERY will get you there.
Prove it on the machine you are reading this on
The same workload runs three times on your device — plain, then with the repeat cache, then with the compressor as well — and the meters sample while it runs. Nothing is sent anywhere. Neither lever reads the model, so this behaves the same whether the request names a model we host or one we have never heard of. The two are separated because they do different things: the cache removes work, the compressor removes bytes and costs a little CPU to do it. Bundling them would hide that.
Other people’s machines
Self-reported · unverifiedThe benchmark runs in the browser, so anything submitted here is whatever that browser chose to send. We cannot verify it without running the work ourselves, which would defeat the point. Treat it as a scoreboard, not as evidence — the evidence is the run you just did yourself.
- Real CPU work, not a simulation: every request is hashed byte by byte. The engine arm is faster only because it skips repeats and processes fewer bytes.
- Your numbers will not match anyone else’s, and that is the point. The milliseconds depend on your CPU, your browser and whatever else your machine is doing — so the run is repeated 3 times and reported as a median with its range, rather than as one draw. What does not vary is the cache hit count and the bytes avoided: those are fixed by the workload seed and come out identical on a phone and on a workstation.
- The cache key includes the model, so two requests for the same text against different models do not share an answer. Raise “distinct models” and watch the hit rate — and the gain — fall.
- Set repeat share to 0% and untick the bloat, and the engine arm will draw level or lose. That case is reachable on purpose; a demo that cannot lose is not a measurement.
- The crossover is real and it is on the slider. Drag “CPU per request” below about 0.25 ms and both levers cost more than they save — a cache lookup and a compressor pass are not free, and at that scale they are more expensive than the work they remove. Measured on one machine: at 0.05 ms per request the cache arm ran 41.7% slower. Above about 1 ms it settles near a quarter faster. A real model call is 100 ms to several seconds, which is why this is worth doing — but the losing end of the range is on the control rather than hidden behind a constant we picked.
- This measures payload work on your device. It is not a comparison against a tuned production serving stack — one of those already does prefix caching and paged attention, and we do not claim their gains as ours.
What is actually measured
Hundreds of modules, 194 of them refusals and 73 improvements over a named alternative, across memory, power, cooling, capacity, siting, GPU scheduling, inference, storage, governance and tenancy. Every one reproduces from a single seeded command.
Three are worked through below, because an example beats a list. They are three of hundreds, not the product — the rest are grouped by who they are for: data centres, AI infrastructure, and regulated industries.
Routing — 32.9% cost cut on a 34,734-request replay
Worked example · costRouting — spend tracks the stakes, not the habit
MEASUREDReplayed against 34,734 real logged requests (9.68 billion tokens), routing cut inference cost from $7,249 to $4,864 while sending every high-stakes call to the premium tier.
The counter-intuitive part: routing is not about making most calls cheaper. It is about the few that aren’t. If you already self-host, the saving goes up — 36.5% API-only, 41.6% with the cheap tail local, 64.2% once your local model carries the balanced tier.
Retrieval quality — better answers reading 12–20× less
Check this one yourselfMemory — the state is the address
PUBLIC DATARetrieval quality on three standard public benchmarks — same corpus, same embeddings, seed 7. You can check these without trusting us at all.
The honest limit on that figure: the reduction above is a coarse-index constant factor. It cuts what you read by a large fixed multiple, and on its own it does not tell you how the curve behaves as the corpus grows.
And over the same sweep it held: from 1,000 to 20,000 documents the read reduction stayed at 10.3–11.3× at every size, and retrieval quality rose rather than decayed — recall@10 from 0.861 to 0.952.
MEASURED 2026-07-10 · BEIR scifact / nfcorpus / fiqa · seed 7 · reproducible offline
Scaling and timing sweep 2026-08-17 · seed 7 / seed 11 · Intel i5-1135G7, 15 W
Finding it — the same speed at a thousand or a million
MEASUREDMost stores get slower as they fill, because finding something means looking through more of it. Ours computes where the answer lives and goes there. We grew the store a thousandfold and timed it.
MEASURED 2026-08-17 ·
bench/directAddressRecall.ts · seed 7 · 1,000 → 1,000,000 items · Intel i5-1135G7, 15 WPrefetch — the one that improves while you use it
MEASUREDIt learns where your work goes next, and gets better at it as it watches. Over 6,000 steps the hit rate more than doubled, with nothing changed but exposure. This module is free, on every tier.
MEASURED 2026-08-17 ·
POST /v1/bench/prefetch · seed 11 · budget 2 of 12Optional, off unless you turn it on, and one file per customer, never pooled — what your traffic teaches the engine is yours and is never mixed with anyone else’s. Local file, no database server, no network call.
MEASURED 2026-08-18 · live engine restart · state reloaded from disk
Ready before the request arrives — 88.4% already prepared
Worked example · answered before you askedReady before the request arrives
MEASUREDMost systems start working when you ask. Ours has usually finished the lookup already, because it prepared for where the work was going. Replayed against real logged traffic, in the order it actually happened:
MEASURED 2026-08-18 ·
bench/warmPath.ts · real logged traffic in arrival order · Intel i5-1135G7, 15 WNoticing work that isn’t yours
MEASUREDA key that leaked. A job pointed at the wrong endpoint. An integration that changed under you. None of it looks wrong line by line — it only looks wrong against what your traffic normally is. So that is what we compare it to, and we never read your prompts to do it.
The honest limits: it is a detector, not a control — it reports, it never blocks. It says nothing until it has seen enough of your traffic to have an opinion, and it says so rather than guessing. And it learns gradual change as normal, by design, so it will not flag a slow drift. No attack appears anywhere in this test and this is not an attack-detection figure.
MEASURED 2026-08-18 ·
bench/anomalyRealTraffic.ts · real cross-workload trafficEnergy — 71.2% less inference load
Energy — the demand side of the problem
ESTIMATEDThe industry is solving data-centre energy with better chips and better cooling. Both are supply-side. Routing is the only lever that reduces the demand itself — and the only one that can be metered per call.
Who it is for
The same decision layer, sized to a different problem in each case. Every one of these runs the engine you can test above — what changes is which constraint is doing the hurting.
Everything above runs in software. That was always the point.
Every figure on this page was produced on ordinary hardware, by software you can install today. None of it needed a new chip — and that is precisely what makes the chip worth building. The architecture is already proven in the only way that counts: it is running, and the results are measured.
Software first, silicon second — in that order, deliberately
IN PROGRESSWhat we will not do here is describe how it works. That is not modesty; patent protection in Europe and China is lost by publishing before filing, and we would rather own it than explain it.
Timothy Harkin
The team and I at Wubbery are really proud of what we have here, and we know that this is a real game changer in many different industries.
A big issue facing the world today is the strain of our ever-growing need for technology — the demand is damaging our planet. I have spent years learning, inventing and developing something that is able to help in a way that not only makes everyone happy, but will be available to everyone eventually too, so we can have a better future together.
Wubbery and I have big plans, and welcome you along for the ride.
Point it at your own workload and watch.
The console ships with the benchmark, the ledger and the shadow replay. If the engine wouldn’t win on your traffic, it will tell you that too.