Nvidia Eyes $100 Billion Quarters, but Is Increasingly Bankrolling the AI Boom

Tuesday, 25 August 2026 at 23:00

Nvidia jaagt op kwartalen van $100 miljard, maar financiert steeds vaker zelf de AI-boom

Nvidia is one of the world’s most important AI companies—and it’s about to test its limits again. The question is: can it pull it off?

Nvidia now faces a far bigger challenge than posting another blowout quarter. The market expects around $92.2 billion in revenue—almost double a year ago—while the company is in the middle of shifting from Blackwell to Vera Rubin and pouring billions into making the infrastructure behind its own AI chips financeable.

That shift changes the core question around Nvidia. It’s no longer just about how many GPUs it can sell. Investors must judge whether demand for AI infrastructure can stand on its own as Nvidia increasingly acts as supplier, investor, buyer, infrastructure partner—and in some cases, guarantor—within the same ecosystem.

That debate matters even more with Nvidia seemingly on track for its first quarter above $100 billion in revenue. The market expects that threshold could be crossed as early as Q3 of fiscal 2027.

Nvidia closes in on its first $100B quarter

Nvidia posted $81.6 billion in revenue for Q1 FY2026, up 85 percent year over year. Data center revenue grew 92 percent to $75.2 billion, according to Nvidia’s official quarterly results.

For Q2, Nvidia guided to $91 billion in revenue, plus or minus 2 percent, implying a range of roughly $89.2 billion to $92.8 billion.

The market, per LSEG consensus cited by Reuters, expects about $92.18 billion—roughly 97 percent year-over-year growth.

That’s extraordinary, but the bar has quietly moved higher than the headline consensus suggests.

Some big banks, per Business Insider, now model around $94–$95 billion. Jefferies is looking to roughly $108 billion the following quarter, and UBS sees $110 billion-plus as possible.

That means a ~$92 billion print could be both a consensus beat and a letdown—because it would still sit inside Nvidia’s own guidance.

The real test likely lies in the outlook for next quarter.

Current LSEG consensus sits around $104.2 billion. If Nvidia guides toward $108–$110 billion, it would suggest Rubin demand is stacking on top of existing Blackwell momentum. A sub-$100 billion outlook would instantly raise questions about how fast growth is cooling.

Data centers now are the Nvidia story

Nvidia’s reliance on AI infrastructure is stark. Visible Alpha expects about $85.7 billion in data center revenue in Q2—roughly 93 percent of total sales.

And growth no longer comes from GPUs alone.

Last quarter Nvidia reported:

$60.4 billion from data center compute, up 77 percent;
$14.8 billion from data center networking, up 199 percent;
about half of data center revenue from hyperscalers;
the other half from AI cloud providers, enterprises, industry, and government projects.

The explosive networking growth is strategically critical.

Nvidia no longer sells just an accelerator. It ships racks, NVLink interconnects, InfiniBand, Spectrum-X Ethernet, system software, storage tech, and full AI data center blueprints.

The more of an AI cluster runs on Nvidia technology, the harder it becomes for customers to swap out only the GPU.

Three direct customers drive 54% of revenue

Behind the record growth is a striking concentration.

Nvidia’s Q1 CFO commentary shows three direct customers accounted individually for 21, 17, and 16 percent of total revenue.

Together, those three made up 54 percent of Nvidia’s sales.

There’s also a second concentration risk. Nvidia says one unnamed AI research and deployment company drove a “meaningful” share of indirect demand through cloud providers.

The company isn’t identified. Without further evidence, it’s not responsible to assert this is OpenAI.

The numbers do show why counting invoice recipients isn’t enough. A server OEM, cloud provider, or distributor may be the direct buyer, while real end demand ultimately comes from a much smaller group of hyperscalers and frontier AI firms.

China isn’t in Nvidia’s current outlook

A notable detail in guidance: Nvidia has included no data center compute revenue from China.

Last quarter Nvidia shipped zero Hopper data center products to China, versus $4.6 billion a year earlier.

So the march toward ~$92 billion is expected without a China rebound.

That makes China a potential upside rather than a pillar of the current outlook. Any loosening of export limits could add revenue, while the base case doesn’t depend on it.

Vera Rubin is far more than a new GPU

The second big question Wednesday is Rubin.

It’s misleading to describe Vera Rubin as just Blackwell’s successor. Nvidia aims to make it a full AI data center platform.

Per Nvidia’s technical overview of Rubin, the platform includes Rubin GPUs, Vera CPUs, NVLink 6, BlueField-4 DPUs, Spectrum-X Ethernet with co-packaged optics, and specialized inference infrastructure.

Nvidia combines these into multiple rack types designed to operate as one giant AI supercomputer.

The company says production is scaling across more than 350 factories in 30 countries, including roughly 150 partners in Taiwan. Product shipments are slated to begin in fall 2026.

That makes the upcoming call critical. Investors should listen closely for the distinction between systems “in production,” systems actually shipped to customers, and systems that can already be recognized as revenue.

Rubin promises massive gains—on Nvidia’s benchmarks

One Rubin GPU is listed at 336 billion transistors, 22 TB/s of HBM4 memory bandwidth, and 3.6 TB/s of NVLink bandwidth.

Versus Blackwell, Nvidia claims up to 5x higher inference performance, 3.5x training performance, and 2.8x memory bandwidth.

At the platform level, Nvidia touts up to 10x higher throughput for agentic AI.

The cost claim is stunning too. Nvidia says Vera Rubin NVL72 can deliver up to 10x lower cost per token than Blackwell NVL72 on certain long-context, reasoning-heavy inference tasks.

There’s an important caveat.

These are Nvidia’s own benchmarks on specific workloads, including a Kimi K2-Thinking test with 32,000 input tokens and 8,000 output tokens under certain latency constraints.

“10x cheaper” does not mean every AI application becomes 10x cheaper on Rubin. It’s a vendor benchmark under specific technical conditions.

Rubin’s arrival could temporarily hit Blackwell

A new Nvidia generation also creates a classic problem.

Customers who know much faster hardware is coming may delay existing orders. Nvidia itself warns in its SEC quarterly filing that new architectures can lead to production delays, lower yields, higher material costs, and additional inventory provisions.

Rubin adds an unusually complex supply chain.

A full system requires not only working GPUs, but also:

HBM4 memory;
advanced packaging;
sufficient manufacturing yields;
liquid cooling;
heavy-duty power;
optical networking;
available data halls;
grid connections;
on-site installation and validation.

A delay in any one element can push out revenue for the entire system.

Reuters also reported, citing Bloomberg, that some Nvidia customers were warned about price hikes above 15 percent for systems delivering in early 2027, including Rubin and Blackwell.

Reuters couldn’t independently confirm, and Nvidia declined to comment.

Higher prices can protect Nvidia’s gross margins—but they also raise the hurdle for customers to earn a return.

Nvidia is trying to mobilize $500B+ in outside capital for AI buildouts

Here’s where Nvidia’s financial pivot comes into focus.

In August, Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build financing platforms that could eventually mobilize more than $500 billion in third-party capital.

That figure is easy to misread.

Nvidia isn’t investing $500 billion itself. There’s no pre-funded vehicle of that size, and the amount is not a guaranteed revenue pipeline for Nvidia.

These are proposed platforms that, over time, could attract over $500 billion of external capital for AI infrastructure.

At announcement, the partnerships still depended on definitive agreements.

Key terms aren’t public: how much each institution will actually commit, at what rates, what collateral is used, which customers get access, and how much credit risk ultimately sits with Nvidia.

Nvidia may support part of the residual value

The standout feature: Nvidia can offer a mechanism on a project-by-project basis to support up to 25 percent of residual value.

In theory, 25 percent of $500 billion would be $125 billion.

That does not mean Nvidia has issued $125 billion in guarantees. It’s an upper bound that would be evaluated per project.

The structure shows Nvidia’s intent.

CEO Jensen Huang wants AI compute to be treated as financeable infrastructure. Nvidia argues GPU systems can be used across customers and models, tuned via CUDA, and even redeployed to other operators.

That could push financiers to view AI compute as an asset that generates multi-year cash flows.

But AI chip residuals are the big unknown

The weak spot: Nvidia’s own pace of obsolescence.

Power hookups, fiber, and buildings can stay useful for decades. AI accelerators cycle much faster.

Rubin underscores the issue.

If Rubin truly delivers up to 10x lower cost per token than Blackwell on select workloads, the follow-on question is what a Blackwell system is worth just a few years later.

Financiers must price something with almost no historical data: the residual value of ultra-expensive AI compute in a market where each new generation can be significantly faster and more efficient.

The OpenAI deal in Ohio makes the risk tangible

The $500 billion platforms are largely intent and future constructs. The OpenAI and SB Energy deal in Ohio is far more concrete.

On August 17, 2026, Nvidia entered multiple agreements with SB Energy for the PORTS Technology Campus project in Pike County.

According to Nvidia’s 8-K filing with the SEC, the initial commitment supports about 4.25 gigawatts of IT load.

Nvidia also has the option to support roughly another 3.8 gigawatts.

The maximum payment obligation under the first commitment is a staggering $105 billion.

Reuters also reported that Nvidia will invest about $1.5 billion in SB Energy.

The first 800 megawatts are slated to come online from 2028. The site could ultimately scale toward eight gigawatts.

OpenAI will be the tenant.

No, Nvidia doesn’t just owe OpenAI $105B

The headline number demands precision.

Nvidia is not lending OpenAI $105 billion, and the amount isn’t immediately due.

Guarantees are activated per lease as capacity becomes ready for use. Leases can run up to twenty years.

A payment by Nvidia essentially becomes relevant if OpenAI becomes insolvent and can’t meet its lease obligations, or otherwise stops paying.

Nvidia would then be on the hook for the gap between a contractually guaranteed minimum value and what SB Energy can recover via a new tenant or a sale.

Nvidia has options: it can, under conditions, assume the lease, find another tenant, or initiate a sale.

OpenAI has also agreed to reimburse Nvidia for amounts Nvidia ultimately pays under the guarantee.

That reduces Nvidia’s legal exposure, but not the credit risk. A scenario severe enough to trigger the guarantee could also make reimbursement harder.

That’s an economic risk assessment—not a damage estimate provided by Nvidia.

The guarantee also locks in future Nvidia demand

The construct matters even more because OpenAI will use Nvidia’s full DSX platform across the guaranteed 4.25 gigawatts, with limited exceptions.

Nvidia is taking financial risk on a data center project that simultaneously secures future demand for Nvidia hardware.

The SEC classifies the agreement as a direct financial obligation or an off-balance sheet arrangement.

This marks a clear step beyond a traditional chip vendor.

Nvidia is helping unlock capital, investing in infrastructure partners, supporting residual value, and in some cases backstopping part of lease obligations—while those facilities are then filled with Nvidia systems.

$105B dwarfs Nvidia’s cash cushion

The nominal size stands out against Nvidia’s balance sheet.

As of April 26, Nvidia reported about $50.3 billion in cash, cash equivalents, and marketable debt securities.

The maximum Ohio guarantee of $105 billion is therefore more than twice as large.

That doesn’t mean Nvidia needs $105 billion on hand tomorrow. Obligations phase in, are conditional, and any recovery from re-leasing or a sale reduces ultimate payments.

It does highlight how far Nvidia is willing to go financially to make AI infrastructure happen.

CoreWeave shows how tightly Nvidia can intertwine with customers

CoreWeave is the clearest live example.

Nvidia is simultaneously CoreWeave’s hardware supplier, shareholder, strategic partner, and a potential buyer of unused cloud capacity.

In September 2025, Nvidia and CoreWeave struck a deal initially valued at $6.3 billion. Under the SEC filing, Nvidia must, under certain conditions, take down remaining data center capacity if CoreWeave can’t fully sell it to others.

The agreement runs through April 2032.

In January 2026, Nvidia then invested another $2 billion in CoreWeave at $87.20 per share. The strategic upside is obvious: every new gigawatt CoreWeave can finance and fill creates potential demand for Nvidia systems.

But it also complicates the economics. Parts of Nvidia’s ecosystem both buy Nvidia hardware and receive Nvidia capital, equity, or capacity backstops.

That doesn’t make revenue automatically artificial. It does mean investors must increasingly separate independent end demand from demand whose financial terms Nvidia helps construct.

CoreWeave also proves real demand is massive

That nuance matters: financial interlinkage doesn’t mean demand isn’t real.

CoreWeave reported roughly $2.58 billion in revenue for Q2 2026. Its revenue backlog stood near $104 billion—contracted future sales not yet recognized as revenue.

The scale of those contracts shows how much future AI capacity customers are trying to lock in.

The problem is shifting. It’s no longer just whether companies want GPUs. The bigger questions are whether data centers can be built on time, whether power is available, and whether the massive investments will produce adequate returns.

Those are precisely the bottlenecks where Nvidia is getting more active.

The AI boom is becoming a financing challenge

The next phase of the AI buildout needs far more than chips.

A data center with tens or hundreds of megawatts of AI compute needs land, transformers, high-voltage connections, cooling, fiber, backup power, buildings, specialized servers—and billions of dollars in hardware.

Multi-gigawatt projects push that even higher.

That’s a fundamental shift from the first wave of generative AI, when clouds could quickly slot more GPUs into existing facilities. The new AI factories require full-scale infrastructure projects with timelines and financing structures that look a lot more like energy and industrial builds.

Nvidia is trying to remove exactly those constraints.

If capital is the bottleneck, Nvidia helps organize financing. If lenders doubt GPU residual values, Nvidia may support part of that value. If a key data center hinges on a single tenant, Nvidia can under conditions backstop part of the lease risk.

That’s rational—so long as underlying AI demand stays strong. It also makes the entire industry more exposed.