INTEGIZER
AI Capacity Per Megawatt
Neural Network Weight Refactoring Between Floating-Point and Integer Domains
← The method itself: INTEGIZER technical page | 📋 Read Technical Abstract
Economics & Business Case
The world's AI capacity is no longer limited by money or by chips. It is limited by electricity — and by grid connections that take four to seven years to obtain.
The largest operators are committing roughly $700 billion to infrastructure this year while waiting in utility queues that capital cannot shorten. Every additional unit of AI output now has to come out of the power a facility already has.
There is substantial headroom in that power. Artificial intelligence does not require floating-point arithmetic — the decimal-number format that dedicated, power-hungry circuitry exists to handle. A trained network can be moved into whole-number arithmetic — which the same silicon executes faster and at lower power, and which billions of cheaper devices can execute when they cannot handle decimals at all.
The INTEGIZER is a patent-pending method for making that move in a structured, reversible and portable way — converting a trained model's weights into the integer domain, retaining the conversion factors as metadata, and restoring the original values on demand.
Everything below is worked in the open: sourced inputs, stated assumptions, and arithmetic any analyst can reproduce in a spreadsheet in ten minutes.
01The Constraint Is Electricity, Not Capital
Two years ago, AI capacity was a procurement problem. Today it is an electricity problem, and electricity does not respond to a better purchase order.
| Constraint | Position as reported |
|---|---|
| Grid connection queues | Utility connection queues in Northern Virginia, Phoenix and Dallas now run four to seven years. A campus joining a Northern Virginia queue in spring 2026 cannot expect utility power before 2030. Sightline Climate data cited by Bloomberg, via SiliconReport, July 2026 |
| Texas grid demand | The Texas grid operator tracks more than 438 gigawatts of large-load connection requests, close to 90 percent of them from data centres. The first realistic transmission plan is not expected until autumn 2027. Utility Dive, via Forbes, June 2026 |
| National backlog | Roughly 2,300 gigawatts of generation and storage capacity sits in United States connection queues — more than the entire installed capacity of the country. Hanwha Data Centers, February 2026 |
| Projects delayed | Around 40 percent of announced AI data centre projects face construction delays caused by power infrastructure rather than chip supply. Nextwaves Insight, June 2026 |
| Regulatory response | In June 2026 the Federal Energy Regulatory Commission issued show-cause orders to six regional grid operators, requiring them to justify or rewrite the rules governing how large loads such as AI data centres connect to the grid. Enterprise DNA, June 2026 |
| Operator workaround | Gas turbines are being built in preference to grid connections because they can be delivered in roughly 18 months against a four-to-seven-year queue. Spheron, June 2026 |
The shape of the workload has also changed permanently. Training a model is a bounded, one-off capital event. Running it — answering queries, which the industry calls inference — is a continuous operating cost that scales with commercial success and never stops.
| Share | What it measures |
|---|---|
| 80 – 90% | Proportion of total AI computing work, and of lifetime computing spend, that goes to running models rather than training them. Analyses compiled by Telnyx, July 2026 |
| ~75% | International Energy Agency estimate for running models as a share of total AI-related energy demand by 2030. Via Voltarione, July 2026 |
| ~85% | Schneider Electric median projection for the same share of computing load by 2028. Cited in arXiv 2506.14551 |
The metric that now governs returns
Not dollars per chip. Useful AI output per megawatt.
Any method that raises output per watt delivers new capacity on a software timetable rather than a grid timetable — measured in months, not half-decades, and without a single additional planning consent.
02Why A Trained Model Can Move To Whole-Number Arithmetic
A neural network's weights measure nothing. They carry no units — no volts, no metres, no dollars — and no single weight means anything on its own.
What a trained model knows is held entirely in the relationships between its weights: which connections are stronger than which others, and by how much. The absolute size of any individual number is an artefact of how it happens to be stored.
This is what makes the conversion sound. Multiply every weight in the model by the same factor and nothing about the model changes. Every ratio is preserved exactly. Every decision boundary sits precisely where it sat before. The network has not been approximated or retrained — it has been re-expressed, the way a measurement is re-expressed when converted between units.
The same idea in familiar terms
Take a restaurant menu priced in dollars and convert every price to cents. Nothing about the menu has changed. The steak is still exactly twice the burger. Every relationship a diner cares about survives untouched — and the conversion reverses perfectly, because you know the factor you used.
The INTEGIZER does this to a trained model: multiply all the weights by a deterministic factor to shift them into whole numbers, keep that factor as metadata alongside them, and divide back whenever the original values are wanted.
The payoff is that whole-number arithmetic is cheaper to perform. Processors handle it with simpler circuitry, less silicon area and less power than the decimal format models are conventionally stored in. Modern graphics processors run it faster on the very same chip. And an enormous population of cheaper hardware can run it when it could not run decimal arithmetic at all.
Where the conversion is exact, and where it trades precision
Whether the conversion is perfect comes down to one thing: whether the destination chip has room for all the digits.
Multiply by 10⁵ 1,234,567,890 → fits comfortably in a 32- or 64-bit whole number
Factor retained 5 → stored as metadata alongside the weights
Divide by 10⁵ 12345.67890 → the original value returns exactly All ten digits survive. The round trip loses nothing in either direction.
Fit to the chip 12,345,678 → only eight digits fit
Factor retained 3 → adjusted to match the digits kept
Divide by 10³ 12345.678 → the last two digits are traded away Whether surplus digits are dropped or rounded is a choice made for the deployment.
The rule is simple: the conversion is exact where the destination chip's range accommodates the full number sequence, and trades precision in proportion to what it cannot hold. Wide destinations — the high-capacity path — convert exactly. Narrow destinations trade digits for speed, memory and reach onto smaller hardware, which is usually the whole point of going there.
At the deployment end of that spectrum, 8-bit whole numbers offer 256 possible values, and well-tuned conversions to that format cost in the region of 1 to 3 percent model accuracy — a figure established across more than 500,000 published evaluations, with some model families more sensitive than others and per-model validation the norm (arXiv 2411.02355; sensitivity per arXiv 2408.15301). The INTEGIZER's contribution is that the retained factor makes this trade recorded and reversible rather than silent and permanent.
03The Arithmetic: A 100-Megawatt Facility
One worked example, every input sourced, every assumption on the table.
100 megawatts is a realistic modern AI campus. The facility is treated as power-capped because that is the real situation: the grid allocation is fixed and cannot be increased on any useful timescale.
Assumptions
| Parameter | Value | Basis |
|---|---|---|
| Grid allocation | 100 MW | Fixed. Additional capacity is unavailable within four to seven years in the major markets. |
| Facility efficiency | 1.15 | Total facility power divided by power reaching the computers. A modern liquid-cooled AI build, between best-in-class fleet figures near 1.09 and the 1.54–1.58 industry average (Uptime Institute surveys, 2024–2025). |
| Power reaching computers | 86.96 MW | 100 ÷ 1.15 |
| Server configuration | 8 processors 10.1 kW per server | Measured: 700 watts per processor with a 1.8× server overhead covering processors, interconnect, memory and power supplies. (Spheron, June 2026) |
| Servers supported | ~8,610 | 86,960 kW ÷ 10.1 kW |
| Processors supported | ~68,880 | 8,610 × 8 |
| Throughput gain from whole-number arithmetic | 1.5× | The low end of the measured range. Derivation in section 04. |
| Electricity price | $0.07 – 0.12 per kWh | Commonly modelled United States commercial range. Large operators frequently secure below-average rates through long-term supply agreements. |
| AI facility build cost | $15 – 20M per MW | Goldman Sachs Research, May 2026, against roughly $10M per MW for conventional hyperscale and $11.3M for standard shell-and-core (JLL 2026 outlook). |
Result A — more output from the same connection
Result A
At a fixed 100-megawatt allocation, running inference in 8-bit whole numbers at a 1.5× throughput gain delivers approximately 50 percent more AI serving capacity from the identical grid connection.
No new connection request. No new substation. No queue position, land, shell, cooling plant or transformer lead time. The capacity arrives on a software deployment schedule.
Result B — the same output from less power
Read the other way round: the original output needs only 66.7 megawatts, freeing 33.3.
| Quantity | Value | Working |
|---|---|---|
| Power freed | 33.3 MW | 100 − (100 ÷ 1.5) |
| Annual energy freed | ~292,000 MWh | 33.3 MW × 8,760 hours |
| Annual value at $0.07/kWh | ~$20.4M | 292,000 × 1,000 × $0.07 |
| Annual value at $0.12/kWh | ~$35.0M | 292,000 × 1,000 × $0.12 |
| Five-year value | ~$102M – $175M | Five years at the above |
Electricity alone. Cooling capital, water and the carrying cost of freed capacity all move in the same direction and are excluded, so this is a floor.
Result C — the cost of buying that capacity instead
Obtaining the same 50 percent increase by building requires roughly 50 megawatts of additional AI-class facility.
| Component | Capital | Basis |
|---|---|---|
| Facility build, 50 MW | $750M – $1.0B | 50 MW × $15–20M per MW |
| Compute hardware, ~4,300 servers | $1.08B – $1.38B | 4,300 servers at roughly $250,000–320,000 each |
| Total capital | ~$1.8B – $2.4B | Sum of the above |
| Grid connection | Cannot be purchased | Four to seven years in the major markets, at any price |
Your own inputs
Every parameter above is a spreadsheet cell. Change the fleet size, the facility efficiency, the power price, the processor generation, or the throughput gain you measure on your own workload, and the model will tell you what it is worth in your environment.
04Where The 1.5× Comes From
Published measurements of 8-bit inference, in full, with the modelling figure set at the bottom of the range.
| Measured | Workload and conditions | Source |
|---|---|---|
| 2.0× peak silicon |
Published chip ratings: the H100 processor is rated at 3,958 trillion whole-number operations per second against 1,979 trillion decimal operations. The B200 is rated 4.5 quadrillion for both. This is a hardware rating, not a deployment result. | Civo, May 2026; Cudo Compute |
| 2.3× | A 7-billion-parameter language model, batch size 16, response latency held under half a second on an H100 processor. | NVIDIA TensorRT-LLM |
| 1.4 – 1.8× | A 70-billion-parameter model on H100 processors at large batch sizes. Inference cost approximately $0.71 per million tokens against $1.19 at the higher-precision baseline. | Spheron, May 2026 |
| 1.6× | A 70-billion-parameter model on four A100 processors, 8-bit whole numbers against the 16-bit baseline under a fixed response-time constraint. | Red Hat / Neural Magic |
| ~1.4× | A prompt-heavy workload on an H100 processor: roughly 40 percent throughput improvement across all batch sizes tested. | SqueezeBits |
The model uses 1.5× — below every production measurement above except the smallest, well below the 2.3× best case, and well below the 2.0× hardware rating. The gain is largest on big models at production batch sizes, where there is enough arithmetic to amortise the conversion overhead. It narrows on small models at low batch sizes: on a compact 256-million-parameter vision model, one published test measured a 6.4 percent slowdown rather than a gain (arXiv 2603.16987). Data-centre inference at scale sits firmly in the first category.
05What The INTEGIZER Adds
Conversion tools exist. What the pending application claims is a conversion that is reversible, portable and self-describing.
Conventional conversion is destructive and one-way: the weights are degraded and the originals cannot be recovered from the converted file. Retarget to different hardware and you go back to the original model and start again. The INTEGIZER keeps the conversion factors with the weights, which changes what the converted asset can do.
🔄 Restoration On Demand
Dividing the whole-number results by the stored factors returns the decimal representation. The converted model stays a live asset rather than a dead end, and the original values remain recoverable within the precision the destination allowed.
🔀 One Conversion, Many Destinations
Because the factors travel with the weights as metadata, a single refactored weight database can be retargeted to different hardware without returning to the original model each time. For an operator running several processor generations, custom accelerators and edge devices, that is an operational saving independent of any throughput figure.
📐 Sized To The Destination
The conversion factor is selected according to the whole-number width the destination hardware supports — so the exact high-capacity path and the compact edge path are two settings of one method rather than two separate pipelines.
📊 Pre-Computed Tables For The Hard Functions
Certain functions inside a network are not simple multiplications and do not scale uniformly. These are handled by pre-calculated lookup tables indexed by whole-number inputs, removing the decimal-arithmetic overhead at runtime rather than merely reducing it.
🛡️ Exact Handling Of Edge Cases
Where a value falls outside the table's range, a decimal fallback is engaged. Because the conversion factor was retained, that specific weight — or the accumulated value entering the function — can be restored to its true original before the calculation runs, so the fallback is exact rather than approximate.
📈 Tables That Improve With Use
Results computed in fallback can be written into the lookup table for later reference, so the same excursion never costs a second decimal calculation. Fallback frequency falls as the system runs, and the same population pass can be performed ahead of deployment for hardware with no decimal capability at all.
The result is that the majority of inference work proceeds through a whole-number arithmetic unit, with the decimal path reserved for the diminishing tail of edge cases.
06The Hardware That Cannot Do Decimal Arithmetic At All
Everything above concerns data centres. There is a second, larger market where the argument is not about speed at all, because the hardware in question cannot perform decimal arithmetic in any format.
These are the chips inside thermostats, doorbell cameras, hearing aids, key fobs, industrial sensors and medical monitors. More than 180 billion ARM-based chips have shipped cumulatively, and ARM Cortex-M cores underpin over 90 percent of the microcontrollers used in small-scale AI work. A Cortex-M has no decimal arithmetic unit. No decimal format reaches it — not 16-bit, not 8-bit, not 4-bit. The only route to intelligence on that silicon is a whole-number route.
The pending application covers this configuration directly: inference performed through a whole-number arithmetic unit doing addition, subtraction and multiplication without a floating-point processing unit present at all. It is not an adaptation of the data-centre case but a distinct configuration in its own right.
Market sizing for this segment varies widely between analysts — roughly $1.36–1.53 billion in 2025/2026, with 2035 projections between $5.5 billion and $9.6 billion. The 30-billion-unit annual shipment figure is the more solid number, and it is ARM's own.
Different economics, different buyers, one method.
07Questions We Are Asked
The industry is moving to 8-bit and 4-bit decimal formats, and the newest chips support them natively. Why move toward whole numbers?
Because most of the world's computing hardware cannot run those formats at all. Compact decimal formats still require decimal circuitry, and the hardware without it outnumbers data-centre processors by orders of magnitude. Three further points:
- On current data-centre silicon, 8-bit whole numbers and 8-bit decimals run at identical peak rates. The choice between them is about handling unusual values, not speed.
- Whole-number-first silicon is widespread beyond the data centre: signal processors, matrix extensions in server processors, tensor units, phone neural engines and the entire microcontroller ecosystem.
- The claimed method is about reversible, retargetable conversion with the factors retained — independent of whether a given destination is whole-number or decimal.
On current data-centre processors the compact decimal formats do handle unusual values more gracefully. That is why both paths are being pursued — and why a method that is indifferent to the destination format carries value either way.
Conversion tools already ship in production. What is left to patent?
Existing tools produce a degraded, destination-specific file. The pending application claims a method producing a restorable, retargetable one: the conversion factor retained as metadata, the factor selected by destination width, pre-computed tables for the functions that do not scale uniformly, and an exact decimal fallback for values outside table range.
What does the conversion cost in model accuracy?
For well-tuned 8-bit whole-number conversion, in the region of 1 to 3 percent — established across more than 500,000 published evaluations. Wider destinations cost proportionally less, and the exact case costs nothing at all. Sensitivity varies by model family, so any deployment begins with validation against the specific models in use.
Why does the model use 1.5×?
Because it understates the measured range. Published production measurements run 1.4× to 1.8×, the best documented case is 2.3×, and the hardware rating is 2.0×. Substitute your own measured figure and the model will return a larger number.
How does an engagement begin?
The INTEGIZER is available for licensing now, under a pending United States patent application. Engagements begin with validation against the licensee's own models and infrastructure — a contained exercise, because the method operates on trained weights that already exist and requires no retraining.
Why has no one else done this?
Decimal arithmetic was inherited from scientific computing, where individual values genuinely do need self-describing scale because a single calculation may span many orders of magnitude. Neural network weights do not have that property — they are dimensionless and cluster within narrow ranges — so the format carries a cost the workload never required.
08Status and Portfolio
| Item | Position |
|---|---|
| INTEGIZER | US application 19/680,833, patent pending. |
| Portfolio | Three granted United States patents — US 12,572,504 B2 (DOORS, file reconciliation through attribute analysis), US 12,368,583 B2 (DSURF, data security using randomized features) and US 11,956,352 B2 (TRIPLE, time randomizing interface protocol language encryption) — together with seven pending applications, the INTEGIZER among them. |
Partner with INTEGIZER Development
Source Index
All figures on this page are attributable. Sources are listed by topic; publication dates are given so that the reader can judge currency.
Power constraint
- SiliconReport, July 2026 — AI datacenter power crunch (Sightline Climate data cited by Bloomberg)
- Forbes, June 2026 — The energy transition just hit the wall (Utility Dive data)
- Hanwha Data Centers, February 2026 — Data center grid limitations
- Nextwaves Insight, June 2026 — Hyperscaler AI capex 2026
- Enterprise DNA, June 2026 — FERC AI data center grid orders
- Spheron, June 2026 — AI data center power constraints 2026 (also the source for node power)
Capex and market context
- Tom's Hardware, April 2026 — Big tech AI spending plans reach $725 billion (FT compilation)
- Futurum Group, February 2026 — AI capex 2026: the $690B infrastructure sprint
- Yahoo Finance, June 2026 — Goldman Sachs $7.6 trillion baseline, 2026–2031
Inference share of load
- Telnyx, July 2026 — AI training vs inference
- Voltarione, July 2026 — European AI data centre electricity benchmark (IEA projection)
- arXiv 2506.14551 — Empirically-calibrated H100 node power models (cites Schneider Electric projection)
Hardware specifications and pricing
- Civo, May 2026 — Comparing NVIDIA B200 and H100
- Cudo Compute — NVIDIA Blackwell architecture: B100, B200, GB200
- Spheron, January 2026 — NVIDIA H100 vs H200
- IntuitionLabs, March 2026 — NVIDIA data center GPU specifications and pricing
- Tom's Hardware, March 2026 — Rack-scale system pricing
Quantisation performance and accuracy
- arXiv 2411.02355 — Give Me BF16 or Give Me Death? Accuracy-Performance Trade-Offs in LLM Quantization (over 500,000 evaluations)
- arXiv 2408.15301 — LLaMA3-70B W8A8 per-channel vulnerability
- arXiv 2603.16987 — The −6.4 percent counterexample: quantisation overhead on a compact VLM
- Red Hat / Neural Magic — LLM Compressor: faster inference with vLLM
- SqueezeBits — vLLM vs TensorRT-LLM: weight/activation quantization
- NVIDIA — Quantization in TensorRT-LLM
- Spheron, May 2026 — FP8 quantization inference performance
Build costs and facility efficiency
- Core Insights Review, June 2026 — AI data center development costs 2026 (summarising JLL 2026 Global Data Center Outlook and Goldman Sachs Research, May 2026)
- Axis Intelligence, June 2026 — Data center construction cost statistics
- The Network Installers, January 2026 — Data center energy consumption statistics
- Semiconductors Insight, May 2026 — Data centre power consumption
Edge and TinyML
- Meta-Intelligence, November 2025 — TinyML engineering benchmarks study (ARM microcontroller estimates)