Gorilla (GRRR), QumulusAI (QMLS) Pass GPU Cost Increases Forward

On August 24 and 25, 2026, Gorilla Technology Group (GRRR) and QumulusAI (QMLS) each told investors on their earnings calls that realized prices for rented AI compute are rising, and that both companies are now passing component cost increases to customers at the moment a contract is signed. Gorilla put the memory increase it is charging forward at 40% to 60% over four to five months [1][2].

What changed on the cost side of renting out GPUs

A merchant AI compute operator does three things: it buys servers packed with GPUs, installs them in a powered and cooled data center, and rents that capacity to customers running training and inference workloads, usually by the GPU hour or by the megawatt. Revenue is typically locked in with take-or-pay contracts, meaning the customer pays the contracted amount whether or not it uses the capacity. The business was built on an assumption that each chip generation gets cheaper and that rents on older cards decline year after year, so operators depreciate equipment over relatively short lives and leave room for falling rents in their pricing.

The direction of that cost line has now changed. Memory, storage, CPUs and GPUs tightened at the same time, which raised the delivered cost of an AI server by a step inside a single contracting cycle, and the tightness sits inside upstream supply agreements rather than in short-term spot swings. Operators found they could not absorb that step within existing price levels, so they began quoting the higher cost to customers at signature. Once one operator does this successfully, competitors have little room to keep absorbing it, because they are all buying from the same component pool.

The prices and contract terms three operators put on the record

Gorilla locked its costs first and then went to price its customers. The company said it sat down again with every single vendor, so everything delivered between now and December runs on fixed prices, and it then quoted the increase to compute customers as a higher upfront charge in order to protect the project economics [1]. The cost of doing so is that Gorilla fixed the customer price in nominal terms for five years. In management's own example, a customer billed $1 today stays at $1 for five years, moving neither down to $0.90 nor up to $1.01 [1].

QumulusAI made the opposite choice on the same question. Its most recent Blackwell contracts generate $18 million to $20 million of annualized revenue per megawatt, against roughly $16 million per megawatt blended across its installed base, and the company calls that gap pricing power while noting that the price per GPU hour has been firming across the market [2]. Because management reads GPU prices as rising rather than falling, it would rather keep some capacity able to reprice than lock all of it to rates set years earlier, which is why the weighted average contract term is held at 2.2 years [2].

The largest merchant AI cloud made the same move. CoreWeave (CRWV) raised list prices by approximately 25% across SKUs in July and said it is passing component price increases through [3]. The specific terms differ across all three companies, but the direction of the quoted price does not.

Contract duration replaces utilization as the variable that decides margin

Utilization — whether the machines are sitting idle — used to be the main thing to watch in this group. With costs and rents both rising, what decides who keeps the increase is how soon a contract comes up for renewal. Capacity already locked at old prices hands the increase to the customer, while capacity that can reprice each year keeps it with the operator. The operating metrics that follow are annualized revenue per megawatt and the share of contracted megawatts that expire or reprice within the next twelve months.

Depreciation is the second line this touches. If rents genuinely stop declining each year, writing equipment off over a short life is conservative, and Gorilla noted that some operators have run A100s in service for seven to nine years and still earn a decent yield on cost [1].

This reading has a clear boundary. Buyers are not behaving uniformly: WhiteFiber (WYFI) said its customers are moving toward longer contract terms because of the same expectation that prices will rise [4]. If buyers broadly lock in long terms, a short-duration operator has no capacity left to reprice. The increase has also not yet appeared in any operator's reported margins, and the pass-through is not clean at the server assembly layer — Super Micro reported non-GAAP gross margin of 17.6% for the June quarter and guided the next quarter to 10.4% to 10.8% [5].

Companies exposed to this change

• Micron Technology (MU): Makes the DRAM and NAND memory that sits on the exact cost line Gorilla renegotiated. It has converted the step into 16 five-year take-or-pay agreements covering calendar 2026 through 2030, with price ceilings set at market prices at signing, covering roughly 20% of its DRAM volume and about a third of its NAND volume [6]. That determines whether operators' cost basis can fall back inside the five-year customer contracts now being signed.
• Applied Digital (APLD): Leases powered, ready-to-populate data center shells, and operators racing to get capacity online often lease existing space rather than build their own, so the urgency lands directly on this layer's rents. The company said rental rates have moved higher over the past six months and that two investment-grade tenants are in negotiation on expansions of roughly 100 MW and 150 MW [7].
• TSS, Inc. (TSSI): Procures and integrates AI racks for operators. Procurement is a low-margin pass-through of the hardware bill of materials, accounting for $18.2 million of $35.1 million in second quarter 2026 revenue [8], so a rising bill of materials inflates that revenue line in dollars while its integration throughput stays gated by the same scarce components.

Sources

[1] Drillr · Gorilla Technology Group (GRRR) · 2026-08-24 · First Half 2026 earnings call

"So there's a higher charge upfront. And we are basically telling, look, this is the risk today. This is the cost of memory. It's gone up 40%, 50%, 60% over the last 4, 5 months. Here is the upgraded cost. And we're making sure that we are intending to protect all of the project economics."

[2] Drillr · QumulusAI (QMLS) · 2026-08-25 · Q2 2026 earnings call

[3] Drillr · CoreWeave (CRWV) · 2026-08-11 · Q2 2026 earnings call

[4] Drillr · WhiteFiber (WYFI) · 2026-08-12 · Q2 2026 earnings call

[5] Drillr · Super Micro · 2026-08-11 · news search event (quarterly gross margin and next-quarter guidance) · https://seekingalpha.com/article/4935624-super-micro-computer-selling-to-someone-other-than-the-hyperscalers-finally-paid

[6] Drillr · Micron Technology (MU) · 2026-06-24 · Fiscal Q3 2026 earnings call

[7] Drillr · Applied Digital (APLD) · 2026-07-27 · Fiscal Q4 2026 earnings call

[8] Drillr · TSS, Inc. (TSSI) · 2026-08-13 · Q2 2026 earnings call

This is only meant to surface industry changes and companies you may have overlooked - it is not a stock recommendation.