Nebius Harlow: The 4.4 MW Head Start
Nebius is installing hardware at Kao Data’s Harlow campus. We trace the 22 MW rollout, test the Q3 revenue claim, and separate evidence from inference.
The public timetable makes Nebius’s Harlow expansion look like a 2027 story. That is only the large half. A paper trail running through Kao Data and its majority owner strongly implies that 4.4 MW sits inside an operational building next door—and Nebius was already installing hardware at the campus by late July.
Four-point-four megawatts is small beside Nebius’s global build. It is large enough to change the near-term question from when will KLON-03 open? to what might switch on before it does?
Full credit to @ai_canaries for surfacing the July interview that sent us back through Kao Data’s disclosures, Infratil’s filings, and Nebius’s deployment commentary. The interview materially strengthens the case, but it does not by itself establish Q3 ARR.
Here is the clean version.
Claim | What the record supports | Status |
|---|---|---|
Nebius contracted for 22 MW at Harlow | Kao disclosed a 10-year agreement for 22 MW | Confirmed |
The deployment spans two data centres | Kao’s founder said so directly | Confirmed |
KLON-03 is 17.6 MW and appears to supply that amount to Nebius | The building size is confirmed; the full customer allocation is inferred from the 22 MW, two-building disclosures | Strong inference |
The remaining 4.4 MW appears to sit in KLON-02 | Subtracting KLON-03’s 17.6 MW from the 22 MW agreement leaves 4.4 MW in the other disclosed building | Strong inference |
The inferred 4.4 MW equals two 2.2 MW technology suites | Kao says KLON-02 has four suites; its engineering partner specifies 2.2 MW per suite | Capacity math confirmed; Nebius tenancy inferred |
Nebius was installing hardware by July 28 | Kao’s CEO said installation was under way at the campus | Confirmed by publication date |
GPU deployment began on July 28 | No source gives a start date or identifies the installed equipment | Unknown |
Could the 4.4 MW contribute to Q3 ARR? | The timing can work at the fast end; activation has not been disclosed | Possible, not proven |
This is not confirmation of a Q3 launch. It is evidence of a potential 2026 deployment path inside a contract most investors can easily file away under 2027.
The 22 MW Was Never One Building
The earliest useful clue predates the public Nebius announcement by seven months.
In Infratil’s November 2025 half-year presentation, the majority owner of Kao Data said KLON-03, together with remaining capacity in KLON-02, would deliver 22 MW. The line now circulating is not a new September construction update. It was published on November 13, 2025, and it did not name the customer.
What it did reveal was the physical shape of a 22 MW block: one building under construction, plus space left in one already standing.
The customer emerged in stages. Infratil’s FY2026 annual report said Kao had secured a 10-year agreement with an international neocloud provider for a 22 MW deployment during the year ended March 31, 2026. On June 8, Kao identified Nebius and confirmed the same terms. Three days later, Kao founder David Bloom wrote that the deployment would run across two data centres.
Put those disclosures beside the campus inventory and the missing capacity has very few places to go.

Kao’s Harlow campus page lists KLON-01 and KLON-02 as operational 8.8 MW facilities. KLON-03 is a 17.6 MW building under development. KLON-04 is planned and much larger.
The subtraction is simple:
22 MW − 17.6 MW = 4.4 MW
KLON-02 was designed with 8.8 MW of IT capacity across four technology suites. Kao’s engineering partner JCA says each technology suite supports 2.2 MW, so two suites make 4.4 MW.
No public floor plan has “Nebius” printed across two numbered rooms. Kao has not identified the precise suites, and neither company has published the building-level allocation in one tidy sentence. That keeps the conclusion in the inference column.
It is a strong inference because three separately published disclosures line up: Infratil names KLON-03 plus remaining KLON-02 capacity; Kao says Nebius spans two buildings; and Kao’s campus design accounts for the 4.4 MW remainder as two 2.2 MW suites.
Harlow Has Two Delivery Clocks
KLON-03 is the long clock.
When Kao announced construction in March 2025, it described KLON-03 as a 17.6 MW liquid-cooled facility designed for dense GPU workloads. Early material pointed to 2026. The more recent, formal timing comes from Infratil’s May 2026 results, which says capacity will come online from early 2027.
“Early 2027” is the disclosed language. Calling it Q1 is reasonable shorthand, but not a company commitment to March 31.
KLON-02 runs on a different clock. The building was fully commissioned and operational in November 2023. Nebius would still need customer-specific electrical and cooling work, racks, GPUs, networking, storage, cluster validation, the cloud software layer, and paying workloads. It would not need to wait for Kao to finish an entire new shell and plant.
That brownfield head start is the thesis. The July interview is what makes it current.
When the Deployment Clock Actually Starts
In a Data Centre Magazine interview published July 28, 2026, Kao CEO Spencer Lamb described Harlow’s opportunity to house 22 MW of NVIDIA GPUs. He then said Kao was delivering against the Nebius contract while the customer installed hardware at the campus.
That confirms physical activity. It does not tell us when the work began.
July 28 is a publication date, not an installation start date. The interview could have been recorded earlier. “Hardware” can also mean racks, network fabric, storage, power distribution, cooling interfaces, GPUs—or several of those at once. The building and megawatt tranche were not named.
The six-to-12-week figure comes from the Q&A on Nebius’s November 11, 2025 earnings webcast, with the wording available in the transcript. Asked about the interval between power connection and revenue, Chief Product and Infrastructure Officer Andrey Korolenko described a general range that begins when power is connected and GPU deployment starts, and ends when the platform is available and generating revenue. Existing-site expansions can move faster.
Both keys matter. An operational building makes connected power plausible; a campus-level comment about hardware makes deployment plausible. Neither condition has been confirmed for a specific 4.4 MW Harlow block.

If July 28 were used only as an illustrative day zero, six weeks lands on September 8 and 12 weeks lands on October 20. September 30 sits in the middle. As of September 4, 2026, even the six-week point is still four days away.
The evidence supports a late-Q3/early-Q4 activation window without establishing either quarter.
Why September Timing Matters
Nebius defines ARR as the final month of the quarter’s AI-cloud revenue multiplied by 12. The definition appears in its Q2 2026 shareholder letter filed with the SEC.
That makes September unusually important.
If a Harlow cluster begins serving customers during September, its September revenue can enter the Q3 exit-ARR calculation—even if it contributes for only part of the month. Reported Q3 revenue, by contrast, records only what the company actually earned in July, August, and September. A late-month activation can therefore show up more visibly in an annualized exit-rate measure than in the quarter’s revenue total.
An October activation affects neither Q3 revenue nor Q3 exit ARR.

There is a more basic boundary. Hardware installation produces neither ARR nor revenue on its own. A hall can be physically ready. Power can be connected. GPUs can be racked. A cluster can be in testing. The cloud platform can be available. A customer can finally consume it. Those are different states.
Nebius’s 2025 annual filing distinguishes contracted, connected, and active capacity for exactly this reason. Our full Nebius model keeps those conversion steps separate as well. A contract gets you a seat at the table. It does not boot a GPU.
The public Harlow record has advanced from contract to installation. It has not publicly advanced to active power, a live customer workload, or recognized revenue.
The 4.4 MW Sensitivity
The 4.4 MW is not transformational beside Nebius’s year-end target of 800 MW to 1 GW of connected power. It can still be financially visible.
The same Q2 shareholder letter gives two useful company-wide reference points. Nebius used roughly $12 million of annual contract value per megawatt for its 2026 base. It said the four landmark Q2 deals averaged approximately $20 million to $25 million of annual contract value per megawatt.
Applied mechanically to 4.4 MW:
Nebius reference | Calculation | Illustrative annual contract value |
|---|---|---|
2026 base: $12M per MW | 4.4 × $12M | $52.8M |
Landmark-deal low: $20M per MW | 4.4 × $20M | $88.0M |
Landmark-deal high: $25M per MW | 4.4 × $25M | $110.0M |

These are ACV sensitivities, not Harlow ARR estimates or forecasts; they assume the full 4.4 MW monetizes at each reference yield.
Nebius has not disclosed Harlow’s GPU generation, usable IT load, customer mix, utilization, pricing, contract type, or realized revenue per megawatt. The 10-year agreement with Kao is an infrastructure input contract. It secures data-centre capacity; it is not a 10-year end-customer revenue contract. Nebius earns its revenue by selling compute and platform services on top.
The Kao lease cannot be mapped directly to Nebius’s customer economics, GPU count, utilization, or Q3 revenue. The sensitivity shows potential scale, not collected revenue.
Why a Small Brownfield Tranche Matters
Harlow is one part of a wider UK build. Nebius said its three new UK deployments would reach a combined 65 MW when fully ramped during 2027. The full 22 MW Harlow agreement is roughly one-third of that footprint.
The inferred 4.4 MW matters less for scale than for execution. If it activates before KLON-03, it would show that Nebius can turn spare capacity inside an operating colocation site into sellable AI infrastructure while a purpose-built facility is still under construction. That is a different test from the larger UK colocation expansion at Longcross or the gigawatt-scale owned build in Pennsylvania.
Kao says Harlow will support Nebius AI Cloud and Token Factory, including inference. That does not identify customers or guarantee utilization, but a live KLON-02 tranche would provide useful evidence of contract-to-revenue execution.
The Case Against the Q3 Call
The bearish reading does not require disputing the 4.4 MW arithmetic. It only requires respecting everything the arithmetic cannot tell us.
Kao’s CEO said installation was happening “at the campus,” not in KLON-02. Equipment could have been staged, shared, or intended for KLON-03. Even an operational shell may need meaningful fit-out before Nebius can draw the reserved IT load. The six-to-12-week range is a general operating heuristic, not Harlow guidance or a service-level agreement.
Then come the last metres of the conversion chain: cluster burn-in, network validation, platform integration, customer onboarding, and utilization. Any one of them can push revenue beyond the physical installation milestone.
None of the Harlow disclosures reviewed for this article ties a named Nebius customer to the site, provides site-specific pricing, or says the inferred 4.4 MW is active.
That is enough to keep Q3 out of the base case today.
It is not enough to dismiss the clue. Public disclosures still place part of the 22 MW agreement in an existing building, and Kao still says Nebius hardware installation is under way.
The Northwise View
Northwise’s Nebius view has never rested on the size of the power pipeline alone. Contracted megawatts only matter when they become connected capacity, usable GPUs, and workloads that pay. Harlow is interesting because it may let us watch that conversion on two timescales at the same campus. KLON-02 offers a small brownfield bridge that could arrive in 2026; KLON-03 brings the larger, likely 17.6 MW Nebius block from early 2027. Treating them as one 22 MW event misses the point.
In our current plan, KLON-03 stays on the disclosed early-2027 schedule, while the inferred 4.4 MW in KLON-02 remains optionality rather than modeled revenue. We would move it into the capacity schedule once Nebius or Kao confirms connected or active megawatts, and into the revenue model only when a live Harlow workload or an attributed commercial contribution closes the loop. Until then, a late-Q3 contribution is upside and Q4 is a labeled scenario. We do not apply the $20 million to $25 million per-megawatt economics from Nebius’s landmark deals without site-specific evidence on hardware, customers, utilization, and contract structure.
Strategically, 4.4 MW will not decide the Nebius valuation. What it can reveal is whether the company has a repeatable way to keep near-term deployments moving while larger facilities are still being built. If KLON-02 goes live first, Harlow becomes a 2026 bridge into a much larger 2027 UK footprint. The value of that bridge is not only the revenue it may produce, but the evidence that Nebius can find capacity, fit it out, bring it online, and put it to work without waiting for every purpose-built AI factory to be finished.
What Would Change the Northwise View
One precise disclosure could settle the building math, but different evidence answers different questions. Kao naming KLON-02 or confirming the handover of two suites would turn the location from inference into fact. Connected or active megawatts would establish that the infrastructure is usable. A UK-region or Token Factory availability update, a production workload tied to Harlow, or management attribution to ARR or recognized revenue would provide the commercial proof needed to change the model.

That is how we will follow Harlow: not by anchoring to July 28, and not by treating every rack as revenue, but by watching the conversion chain in order. First comes the physical allocation, then connected power and available compute, and finally a paying workload. Each step strengthens the case, but only the last one turns an installation update into financial evidence.
Until that chain advances, KLON-02 remains credible 2026 optionality and KLON-03 remains the larger early-2027 delivery. The 4.4 MW itself may be modest, but the process it tests sits at the center of the NBIS thesis: whether Nebius can repeatedly turn secured power into reliable, occupied compute at the cadence its plan requires. If Harlow reaches customers before KLON-03 opens, the small building will have told us something important about the big plan—and about how Nebius intends to build the rest of it.
Reader discussion
Discuss the research
Create a Free account to join the discussion with a display name.
Create a Free Account




No comments yet. Start a thoughtful discussion.