The Great Disintermediation: How Nvidia Squeezes Its Partners to Seize Control of the Supply & Demand

As Nvidia is fighting the law of large numbers in order to continue the “beat and raise” tradition, which they intend to maintain for as long as possible, otherwise “the whole world would have fallen apart,” Nvidia has, over the last two quarters, taken near complete control of the GPU supply and demand market. And when you need full control over supply and demand, you cannot have intermediaries, namely Nvidia’s partners, sitting in the middle. You have to sell directly and communicate directly with end customers. Therefore, starting last quarter, at least the first time they disclosed it, and possibly earlier, knowing Nvidia, the company decided to execute the great disintermediation and squeeze its partners out of key channels as part of seizing control over supply and demand.

But before getting to the main point, let me provide some background on Nvidia, its partners, and their relationships.

Historically, Nvidia designed the GPU and reference board, which it then licensed to add-in board partners such as Asus, Gigabyte, PNY, and ZOTAC, who finalized the products with custom PCBs, cooling solutions, and distribution to end customers.

In the AI and data center market, Nvidia works with OEMs and system builders such as HPE, Dell, Gigabyte, and SMCI, which integrate Nvidia GPUs into servers and racks based on Nvidia’s reference platforms and system architectures, with far less room for differentiation.

Being an Nvidia partner is hard. Very hard. Why? Let’s go back to EVGA’s exit from the GPU market in 2022. When EVGA was asked about its exit, it cited lack of supply, the level of financial control Nvidia exercised over board partners, low margins, and direct undercutting competition from Nvidia’s Founders Edition cards, which made the GPU business unprofitable and a poor use of time.

Exactly. Nvidia has complete control over everything, including pricing, leaving its partners with low single digit percentage profits, if any at all. And if Nvidia decided to launch a Founders Edition card, it could undercut its partners, causing those partners to effectively lose money.

As proof, look at the net profit margins of the major AI server builders: HPE, Dell, Gigabyte, and SMCI. Many assumed Nvidia’s partners would benefit massively from the AI bubble and the FOMO around AI servers. Instead, we see razor thin margins in the low single digits. SMCI has faced accounting issues and has repeatedly had to raise funds.

And Nvidia? 55% NET PROFIT MARGIN, effectively capturing almost all the value and leaving only crumbs to its partners

As time passed and GPU supply increased, with competition from AMD and Google TPUs emerging, Nvidia began to face growing margin pressure from its customers. The question then became how to protect margins while remaining competitive and continuing to sell large volumes of GPUs. The answer was to squeeze its partners by moving closer to the end customer and selling directly.

Looking at Nvidia’s revenue concentration disclosures from 2023 to 2025, several important developments stand out in the most recent quarter:

First, Nvidia disclosed for the first time that it is selling directly to CSPs and hyperscalers rather than relying solely on its traditional partners. It is possible that this practice has existed for some time and is only now being disclosed, which could raise questions about earlier reporting. For the purpose of this analysis, I will assume Nvidia’s disclosure is accurate and that this shift occurred for the first time during this quarter.

Second, in Q3 2025, Nvidia, for the first time, introduced “neocloud builders” and “hyperscale” companies as explicit categories. Nvidia and its partners have been serving, and in some cases investing in, these companies since at least 2023. The timing of this new categorization raises questions, one possible explanation being related to the next point.

Third, Nvidia disclosed that some CSPs and hyperscale companies are now direct customers, while others remain indirect customers through partners.

In effect, Nvidia is telling us that it is now communicating with and selling directly to CSPs and hyperscalers such as Microsoft, Oracle, CoreWeave, Amazon, and others, reducing the role of OEMs and system integrators like SMCI, Dell, and HPE. This allows Nvidia to capture margin that previously accrued to those OEMs.

But margin capture is only one reason of squeezing partners.

This shift implies several additional effects.

First, Nvidia gains direct control over allocation, meaning who gets what and when. This is critical when supply remains constrained and demand must be actively managed.

Second, Nvidia gains direct visibility into hyperscaler buying behavior instead of relying on secondhand information from intermediaries.

Third, Nvidia can recognize revenue at the point of sale to the hyperscaler rather than waiting for a partner’s resale cycle, which affects revenue timing.

Fourth, Nvidia can accelerate revenue recognition compared to a model in which it must wait for partners to sell through inventory.

Fifth, Nvidia can use bill-and-hold arrangements more aggressively, which is particularly important for Blackwell, as I discussed in my article on the Blackwell story. Bill and hold allows Nvidia to recognize revenue before physical delivery. Many customers want to lock in supply even if their data centers are not yet ready from a power or cooling perspective. They pay Nvidia to manufacture the racks and hold them. Under specific accounting rules, Nvidia can recognize revenue immediately upon completion of manufacturing, even if the racks remain in Nvidia’s warehouse for months. Direct customer relationships make these arrangements far easier to execute than when intermediaries are involved. Additionally, in a very strong quarter, Nvidia could choose to delay recognizing some revenue by asserting that systems are not yet ready, effectively saving revenue for a future quarter when results might otherwise fall short.

Sixth, with direct control over customers and supply, Nvidia can exert pressure to meet quarterly targets by pushing partners or customers to commit to capacity earlier than needed, while offering mechanisms such as buybacks or capacity rentals. Examples include arrangements involving companies like CoreWeave or Lambda.

Some may recall similarities to the Sunbeam “side letter” practices during the dot-com era. In the Sunbeam scandal, the CEO sold grills to retailers during winter using bill-and-hold arrangements, while secretly issuing side letters allowing unsold inventory to be returned months later. This made the sales economically meaningless, even though revenue was booked.

Nvidia has effectively declared war on the law of large numbers and intends to continue the “beat and raise” tradition for as long as possible, using every available means to do so.