What was announced, and what was left out
On October 21, 2026, DataCenterDynamics reported that Vultr will deploy a 20 MW cluster of AMD Instinct MI355X GPUs in Chennai, which the company calls the largest AMD Instinct deployment in India. The announcement named the megawatts. It did not name the data center, the operator, or the number of GPUs.
That missing number is the interesting part, because it is recoverable. Vultr’s own earlier deployment gives the conversion factor.
The conversion factor Vultr already published
The Springfield, Ohio cluster announced in December 2025 is a 50 MW facility with 24,000 MI355X GPUs, per Reuters, BusinessWire and DCD. That is 2,083 watts per GPU at the facility level.
AMD’s published figure for the chip is 1,400 W TDP. InferenceX publishes 2.09 kW all-in per chip, which is the board plus the host CPU, the network interface and the share of cooling. The Springfield number lands almost exactly on that figure: 24,000 chips at 2.09 kW is 50.2 MW. So the 50 MW headline is a facility figure, not a board figure.
Apply the same factor to Chennai and the count falls out: 20 MW divided by 2,083 W is about 9,600 GPUs. If you read the 20 MW as IT load only, at 1,400 W per board, you get 14,286. The first number is the one consistent with Vultr’s own published pair, so 9,600 is the defensible read.
Vultr has not confirmed this. It is arithmetic on a published ratio, not a disclosure.
Why the facility figure is the one that matters
The gap between 1,400 W and 2,090 W is 49 percent. That gap is the host CPU, the NIC, the power delivery losses and the cooling. On a single card it looks like a rounding detail. Across 9,600 cards it is the difference between 13.4 MW and 20 MW, and between roughly $11 million and $17 million a year in electricity at Tamil Nadu rates.
Any energy-per-token claim that uses board TDP is understating the bill by about half. SemiAnalysis uses the all-in figure for its own energy-per-token math on the MI355X page, which is the correct convention.
Energy per token, from published throughput
Throughput per GPU on MI355X is measured continuously on InferenceX. The GLM-5 FP8 SGLang recipe at ISL 8192 / OSL 1024 on a single node produced these figures as of the May 20, 2026 snapshot:
| Concurrency | tok/s per GPU | J/token, board only | J/token, all-in |
|---|---|---|---|
| 4 | 625.5 | 2.24 | 3.34 |
| 8 | 911.7 | 1.54 | 2.29 |
| 16 | 1,208.1 | 1.16 | 1.73 |
| 32 | 1,707.4 | 0.82 | 1.22 |
| 64 | 1,895.0 | 0.74 | 1.10 |
| 128 | 1,911.7 | 0.73 | 1.09 |
These are divisions of published power figures by measured throughput, not independent measurements. The shape of the curve is the point: energy per token falls by a factor of three between concurrency 4 and concurrency 128. Nothing about the chip changed between those rows. Only the batch size did.
What a single user actually costs
At batch size 1 you read the model weights for every token you generate. That is why the low-concurrency rows are so expensive per token. I measured this on my own setup to see how far the single-user case sits from the data center case.
The serving box is a Ryzen 9 5900X with an RTX 3090, running Strata. The request was sent from a different machine, a Ryzen 7 2700X, so the wall-clock figures include the network round trip. Power was sampled over SSH during each request. I ran the same 729-token prompt four times. Decode energy per token came out between 3.16 and 3.46 J across the four runs — a 9 percent spread, which is the point of running it four times.
| Stage | Throughput | Power drawn | J/token |
|---|---|---|---|
| Idle, model resident, 0% utilization | — | 16.5 W | — |
| Prefill, 729 tokens | 517 to 654 tok/s | not separable, see below | — |
| Decode, 1,200 tokens | 99.4 to 105.8 tok/s engine, 97.9 to 100.9 tok/s wall-clock | 330.9 to 343.6 W mean | 3.16 to 3.46 |
3.16 to 3.46 J per token on a 2020 consumer card. The MI355X at concurrency 4 is 3.34 J all-in. Those are effectively the same number, and that is the finding: at batch size 1, a flagship data center accelerator is not meaningfully more efficient per token than a six-year-old consumer GPU. It is about six times faster, which is a different thing entirely.
Two things about the measurement are worth stating plainly. The idle figure is 16.5 W at 0 percent utilization. A reading taken right after a request finishes shows 257 W or 288 W at 0 percent utilization, because the card keeps drawing while the driver settles, so that residual is about sixteen times the true idle baseline and is not a valid idle number. And prefill power is not separable here: the window is 1.1 to 1.4 seconds and the samples inside it read between 17 W and 222 W, so any per-token energy for prefill is noise. Decode is the phase that is stable, and it is the phase that costs money.
Per byte moved, the gap does show. Energy per byte is board power divided by memory bandwidth, assuming the bus is saturated during decode. On my card that is about 335 W over 936 GB/s, which is 0.36 J per GB. On the MI355X it is 1,400 W over 8,000 GB/s, which is 0.175 J per GB. HBM3e is roughly twice as efficient per byte as GDDR6X, and that ratio is the only hardware difference between the two rows above.
Both figures are my own arithmetic. The 936 GB/s is TechPowerUp’s published spec for the 3090, not something I measured, and the 8,000 GB/s is AMD’s published figure. The wattage is the number I sampled.
The tariff is a bigger lever than the chip
Energy per token is fixed by the hardware and the batch size. The price per kilowatt-hour is fixed by the state and by which tariff category the distribution company assigns to the connection.
FlexiCloud’s October 2026 compilation puts Indian grid power for large users between about ₹6.2 per kWh in Andhra Pradesh and ₹10.9 in Maharashtra’s high-tension industrial category. Tamil Nadu sits at about ₹9.0 on FY2025-26 rates. At the FBIL reference rate of ₹95.25 per US dollar, that is $0.094 per kWh.
At $0.094 per kWh, the 1.22 J per token at concurrency 32 costs $0.032 per million tokens in electricity. At Andhra Pradesh’s rate it is $0.022. At Maharashtra’s high-tension industrial rate it is $0.039. The full quoted cost on InferenceX for that configuration is $0.24 per million tokens, so electricity is about 13 percent of it.
That 13 percent is not the interesting number. The interesting number is the tariff category swing inside one state: Maharashtra’s high-tension commercial supply averages ₹16.74 per kWh against ₹10.88 for high-tension industry in FY2026-27, a 54 percent difference for the same electrons. Across states the spread is wider still, 75 percent between Andhra Pradesh and Maharashtra. Choosing how a connection is billed moves the energy cost more than choosing a different GPU does, and it costs nothing to ask for it in writing.
Across the derived 9,600-GPU fleet at 2.09 kW per chip, that is 20,064 kW and about 175 GWh a year. At Tamil Nadu rates that is roughly $16.6 million. At Andhra Pradesh rates, $11.4 million. The difference between two states is $5.2 million a year on the same hardware running the same workload.
Building is cheaper than power in India. Cushman & Wakefield puts a data centre build at about US$7.4m per MW against US$15.2m in the USA. Colocation rent in Mumbai is US$73 to US$105 per kW a month against US$190 to US$235 in Northern Virginia. On 20 MW that is a $156 million difference in capital, and between $28 million and $31 million a year in rent depending on which end of each range you take. The Chennai decision is not primarily about the chip.
What this means for a local build
The single-user case is the one where energy per token is worst, and it is worst for a structural reason rather than a hardware one. Batching is the lever. A provider that serves you at concurrency 4 is spending 3.34 J per token on your request. The same hardware at concurrency 32 spends 1.22 J. If your workload can be batched, the price per token should follow the curve, and the places where it does not are where the provider is holding concurrency low for latency reasons.
For a local build, the practical consequence is that power draw is not a proxy for efficiency. My 3090 draws about 335 W and produces around 100 tokens a second. A board rated at 1,400 W produces 1,707 tokens a second under load. The per-token cost is what matters, and you can measure yours with a plug-in meter and a stopwatch.
The metering point is worth taking seriously because it is the cheapest instrument on the site. A plug-in watt meter gives you the draw during decode, which is the number that turns into J per token when you divide by the throughput the API reports. That is the same measurement I ran, just with a meter instead of a remote query. The tokens per watt arithmetic is where that number becomes a decision, and a power cap changes it in ways that are not obvious from the throughput curve alone.
The other half of the decision is memory. Energy per token is dominated by bytes moved, and bytes moved is a memory question before it is a compute question. The bandwidth per dollar framing is the one that survives contact with a bill, because a card that moves fewer bytes per second costs more energy per token at the same batch size. The DRAM price squeeze behind HBM supply pricing is what makes the 288 GB tier expensive to build even when the silicon is available.
What to buy if you are measuring rather than guessing
The practical version of this article is a meter and a stopwatch. The P3 International Kill-A-Watt reads watts, kilowatt-hours and cost, and it is the instrument that turns a throughput number into an energy-per-token number. It is the cheapest thing on this page and the only one that produces data you can check.
P3 International Kill-A-Watt Electric Usage Monitor — about $34.99, in stock.
If you want the reading to persist rather than be watched, a networked monitor logs it. The Emporia plug is device-level only, which is enough for a single workstation on a dedicated outlet. It is about $12.
Emporia Energy Monitoring Smart Plug, 15A — about $11.99, in stock.
The card I measured on is a renewed Dell M8HMD RTX 3090, which is the cheapest way to get 24 GB of GDDR6X and a 936 GB/s bus. It is about $1,800 at the time of writing, with three units listed. The median lowest used price on eBay was $1,199 in October 2026, 22 percent above its 12-month average of $983, so the renewed listings sit about 50 percent above the used market. That premium is the warranty, not the silicon.
Dell M8HMD NVIDIA GeForce RTX 3090 24GB (Renewed) — about $1,799.99, three listed.
The number that is not disclosed
Vultr announced 20 MW and left the GPU count out. The count is 9,600 by the company’s own published ratio. Nobody has confirmed it, and the facility it lands in is also unnamed, with Digital Realty, Equinix, Sify, CtrlS, NTT Data, CapitaLand and Iron Mountain operating in the area.
What is not in dispute is the shape of the economics. Energy per token falls threefold between batch size 4 and batch size 128 on the same silicon. The price per kilowatt-hour varies by 75 percent between two Indian states, and the tariff category varies by more than that within one state. A deployment in Chennai is a bet on the second number being smaller than the first, and on the building costing half what the same building costs in Virginia.

