NVIDIA's AI rack, the NVL72, is starting to resemble a capital good more like a power plant or an aircraft than a semiconductor product. On August 28, 2026, TrendForce forecast that NVL72 production value will exceed $710 billion in 2027, a 214% increase from the prior year. The forecast covers GB300 as well as the Vera Rubin-series VR200 and VR300. Because shipment growth is expected to stay above 50%, most of the increase stems from generational transition and rising per-unit value. What's more, NVIDIA has gone so far as to guarantee land, power, and buildings, and to arrange third-party capital, so that customers can afford to deploy this expensive equipment.
Behind the $710 Billion Figure: Price, Not Volume
According to TrendForce's forecast, GB300 will be the mainstay product in 2026, with demand continuing through the first half of 2027. After that, production of the Vera Rubin-series VR200 will ramp up, accounting for the largest share of 2027 shipments. The newer Rubin Ultra-series VR300, meanwhile, is expected to represent only a small share that year, since its HBM configuration and server design have not yet been finalized.
Using TrendForce's two rounded figures as given, 2026 NVL72 production value comes to roughly $226.1 billion. If shipment growth is assumed to be 50%, the average production value per rack would need to rise by roughly 2.09 times to reach a 214% increase. Since actual shipment growth is described as "above 50%," the required multiplier would be somewhat below 2.09. Even so, the underlying relationship holds: product mix and unit price contribute more than volume growth. This is why TrendForce expects Vera Rubin's average selling price (ASP) to run roughly twice that of GB300, and why it factors in rising wafer and HBM prices.
That said, the $710 billion figure is not a confirmed order total. While TrendForce's public statement cites $710 billion for 2027 and 214% year-over-year growth, it does not disclose rack shipment volumes, per-model ASPs, 2026 production value by model, or the scope of components counted—making it impossible to independently reproduce the forecast from public information alone. The paid report may contain more detailed inputs, but what can be verified from the free announcement is only the basic structure: volume and price combining to produce a 3.14x multiple.
Conflating this production value with NVIDIA's own revenue risks misreading the size of the market. TrendForce's figure covers entire rack systems assembled by ODMs, which is a different scope from NVIDIA's standalone GPU, CPU, and networking sales. NVIDIA's fiscal Q2 FY2027 revenue was $96.221 billion, of which data center revenue was $89 billion—over 90% of the total. That figure illustrates NVIDIA's business concentration, but it cannot be directly compared with the $710 billion figure.
Same 72 GPUs, a Very Different Rack
Both the GB300 NVL72 and the Vera Rubin NVL72 integrate 72 GPUs and 36 CPUs into a single rack. The outward unit count is identical, but the GPU generation and CPU change, and memory and intra-rack communication specifications shift substantially. Comparing unit counts alone fails to capture the rise in rack value that comes with generational change.
| Published Spec | GB300 NVL72 | Vera Rubin NVL72 |
|---|---|---|
| GPU / CPU | 72x Blackwell Ultra / 36x Grace | 72x Rubin / 36x Vera |
| GPU Memory | 20TB, HBM3E | 20.7TB, HBM4 |
| GPU Memory Bandwidth | Up to 576TB/s | Up to 1,580TB/s |
| NVLink Bandwidth | 130TB/s | 260TB/s |
According to published specifications, total NVL72 GPU memory bandwidth increases roughly 2.74x, from a maximum of 576TB/s for GB300 to a maximum of 1,580TB/s for Vera Rubin, while NVLink bandwidth doubles from 130TB/s to 260TB/s. The increase in GPU memory capacity is modest by comparison. The value of this generational shift lies less in "holding more" than in "moving in and out of roughly the same amount of memory far faster." At minimum, the physical basis for the per-unit price increase that can be confirmed from published specs lies in bandwidth and intra-rack communication improvements rather than capacity growth.
NVIDIA itself has explained that revenue opportunity per gigawatt has expanded from roughly $18 billion with Hopper to roughly $25 billion with Blackwell and roughly $40 billion with Vera Rubin. This is not realized revenue but an opportunity estimate assuming the company captures sales not just of GPUs but also CPUs, NVLink, InfiniBand or Ethernet, and inference accelerators. Rising rack prices are not driven by GPU price increases alone—the scope of components and software entering the data center is itself expanding.
Note that Vera Rubin's product page labels these figures as preliminary specifications. Maximum bandwidth is neither the sustained performance an actual application can achieve nor the cost per token. Published specs indicate a physical reason for rising unit prices, but whether that price is justified will be determined by utilization rates and real-world workloads.
Even With HBM4 in Mass Production, Supply Remains Tight
Mass production of Vera Rubin has already begun. NVIDIA announced the start of mass production in its August 26 earnings report and projected that Vera Rubin will account for roughly 20% of data center revenue in fiscal Q3 FY2027. Micron, as of June, was already shipping HBM4 in volume to a major customer's platform. The next generation, HBM4E, is slated to enter mass production in 2027.
The start of mass production and the availability of supply are not the same thing. While NVIDIA has built inventory up to roughly $32 billion in preparation for the Vera Rubin rollout, it expects supply to remain a bottleneck at least through the end of its fiscal 2028, i.e., through the end of January 2028. Memory price increases have also exceeded expectations, with further increases expected into the following year, according to the company. The HBM price increases TrendForce has built into its forecast push production value higher, but for NVIDIA they constrain both cost and shipment volume.
This dual nature widens the range of uncertainty in the 2027 forecast. Rising HBM prices push up rack production value, but if supply bottlenecks persist, shipment volumes will not grow at the same pace even as prices rise. Furthermore, TrendForce's view that VR300 will remain small-scale reflects not weak demand but the fact that its HBM configuration and server design remain undetermined. With unit price, supply volume, and product mix all moving simultaneously, there is no single path by which the $710 billion figure gets realized.
A Falling Share for the Big Five Signals Broader Demand
Buyers are changing too. TrendForce forecasts that the share of NVL72 procurement accounted for by North America's top five cloud service providers (CSPs) will fall from roughly 70% in 2025 to 60% in 2026. This is not a forecast that the big five will buy less in absolute terms. Rather, it reflects an expanding denominator, as demand from Tesla, xAI, SpaceX, CoreWeave, and others grows quickly.
NVIDIA's own results show the same trend. Of the $89 billion in data center revenue in fiscal Q2 FY2027, $49 billion came from hyperscalers. ACIE—which combines GPU-specialized emerging clouds, industrial, and enterprise customers—reached $40 billion. NVIDIA expects neocloud installed capacity to grow from roughly 3GW at the end of 2025 to 8GW by the end of 2026.
As the customer base broadens, dependence on the capital spending plans of a handful of companies declines. On the other hand, disparities widen among buyers in terms of financial strength, equipment utilization, and resale capability. For large purchases by investment-grade customers, NVIDIA sometimes offers payment terms longer than usual—ranging from 90 days up to a full year. The seller is no longer simply handing over the product; it has begun supporting customers through the time it takes to sell and monetize computing capacity.
Chip Companies Are Now Underwriting Demand Itself
On August 10, 2026, NVIDIA announced a financing platform with six major asset managers and investment banks, including Apollo, aimed at drawing in more than $500 billion in long-term third-party capital for AI computing infrastructure. The goal is to turn expensive GPU racks into assets investors can hold, while enabling cloud operators and AI companies to use them under long-term contracts. The cost of capital paid by customers has now become part of the competitive landscape for semiconductors.
The PORTS-Pike Technology Campus in Ohio illustrates the scale involved. SB Energy will build, own, and operate the facility, with OpenAI planning to lease roughly 8GW under a 20-year contract. NVIDIA has invested $1.5 billion in SB Energy and put in place guarantees of up to $105 billion for the initial 4.25GW. The guarantee takes effect once the nine-phase facility becomes operational and decreases as OpenAI pays lease fees. It also covers only a defined portion of lease fees and power costs—not the site's total costs or all of OpenAI's obligations.
A similar structure is spreading to custom silicon as well. Broadcom, Apollo, and Blackstone are backing more than 1GW of computing infrastructure for Anthropic through an initial $35 billion AI XPV Platform. According to Broadcom's SEC filings, the company had Apollo assume AI rack purchase agreements and computing capacity leases, and guaranteed the customer's five-year lease obligations.
These kinds of guarantees can pull demand forward, but they don't eliminate risk. NVIDIA has disclosed that it has agreements to purchase capacity itself if a supported AI cloud cannot sell its contracted capacity to a third party. PORTS-Pike likewise requires infrastructure and permits, and won't move forward unless environmental review and financing come together. The existence of guarantees and buy-back agreements signals that demand risk—the possibility that usage revenue falls short of projections—still remains. However, who ultimately bears losses, and to what extent, cannot be determined from public materials alone.
The credibility of the $710 billion forecast can be gauged well before 2027 shipment figures are published. Watch whether Vera Rubin reaches 20% of data center revenue in the coming quarters, whether NVIDIA can narrow the supply shortfall by the end of January 2028 while absorbing memory price increases, and whether PORTS-Pike begins phased operations starting in 2028. Once high-priced racks can sustain 20-year contracts on operating revenue alone, AI computing infrastructure will no longer depend on chip-company guarantees—it will have become a self-sustaining infrastructure asset.