With 81.6 billion in single-quarter revenue, can it really hold up to 5.45 trillion?

NEXT PICK · Market Insights

With 81.6 billion in single-quarter revenue, can it really hold up to 5.45 trillion?

Monday, August 17, 2026

A market cap of $5.45 trillion paired with a price-to-earnings ratio of 34.46 times is paradoxical in itself—the world's most expensive company carries a valuation multiple that even many consumer stocks can't match.

Even more dazzling is the rhythm

The most recent full fiscal year reported revenue of $215.9 billion, a year-on-year increase of 65%, while the recently concluded quarter alone reached $81.62 billion, an 85% year-over-year increase. The growth rate is not slowing but accelerating.

However, the stock price remained flat at $225.01, down 0.07% for the day, while during the session, $4.9 million in large call orders were held above 2.6 times the bearish side. Nine days later, the quarterly report on August 26 will put the answer on the table.

Latest price $225.01 ▼ -0.07%

Price-to-earnings ratio 34.46 times — Forward-looking is about 24 times higher

Last quarter's revenue 81.62 billion USD ▲ Year-on-year +85%

A company that no longer sells graphics cards is selling an entire AI factory

If NVIDIA is still regarded as a graphics card manufacturer, you completely misunderstand the company's revenue structure: of the most recent full fiscal year's $215.9 billion revenue, data center business contributed $193.7 billion, a 68% year-on-year increase, while gaming graphics cards have dropped to just over 10%.

The recently disclosed quarter was even more extreme: data center sales reached $75.25 billion, a 92% year-over-year increase, with demand for network products like InfiniBand and NVLink doubling—indicating that customers are not buying a single chip, but a whole set of computing power clusters with pre-connected cables and finely tuned parameters.

So what makes customers avoid comparing prices? The answer isn't in silicon wafers, but in CUDA, which was launched in 2006. The muscle memory of millions of developers worldwide, native support for TensorFlow and PyTorch, and the entire optimization library of cuDNN and TensorRT together form an invisible foundation.

It takes a deep learning engineer about half a year to a year to master CUDA, and migrating a well-functioning model to AMD's ROCm or Intel's oneAPI means rewriting the kernel, re-tuning, and retaking stability risks—a cost that is outrageous for labs chasing model iterations.

Even more severe is vertical integration. GPU, Grace CPU, BlueField DPU, NVLink and NVSwitch interconnection, InfiniBand and Spectrum-X networks—only one company worldwide has developed and fully developed the system simultaneously. The GB200 NVL72 cabinet uses NVLink to string together 72 GPUs into one machine.

Rather than selling accelerator cards, NVIDIA is selling a turnkey power plant: competitors can build cheaper generators but cannot build entire power grids.

A 34x P/E ratio is not expensive, but the inversion between GAAP and non-GAAP is glaring

First, break an intuition

Based on a $225.01 share price and a rolling EPS of $6.53, the price-to-earnings ratio is 34.46, while the forward P/E ratio is just over 24—looking backwards, the market consensus for earnings per share for this fiscal year has already surpassed $9.

The real value lies elsewhere. A market capitalization of $5.45 trillion corresponds to $96.6 billion in free cash flow in the most recent full fiscal year, with a free cash flow yield of less than 1.8%; The price-to-sales ratio on rolling revenue is also above 21 times, and all these anchors point to the pricing premise that "growth must be delivered."

So why do the rolling P/E ratio and the forward P/E ratio differ by ten multiples? Because the denominator is rapidly expanding, and this is precisely the sharpest argument for bulls—and also the point bears fear most—once the growth rate slides downward, valuation compression will happen faster than earnings decline.

The quality of cash flow is beyond reproach. In the most recent full fiscal year, operating profit was $130.4 billion, GAAP net profit was $120.1 billion, gross margin was 71.1%, free cash flow was $96.6 billion, and for the full year, $41.1 billion was returned to shareholders, with a repurchase authorization balance of $58.5 billion.

But there is one detail worth the reader's pause

Last quarter, GAAP earnings per share were $2.39, which was actually higher than the non-GAAP $1.87. This kind of inversion usually comes from non-operating income, meaning that of that astonishing 71.5% quarterly net profit margin, one part wasn't earned from chip sales.

This makes the judgment an awkward position

It's reasonably expensive, but the margin of safety is almost squeezed dry. It's not bubble pricing, but perfect pricing—the biggest fear isn't bad news, but news that is "not good enough."

Can a company with annual revenue of over 300 billion still be called a growth stock?

Based on last quarter's annualized $81.62 billion, NVIDIA's revenue has already exceeded $320 billion, ranking first in market value globally. Its customers are Tier-1 cloud providers like Amazon, Google, Microsoft, and Meta, and it has not invested a single cent on the pace of startup fundraising.

The product roadmap has long moved beyond the stage of "whether the market will accept it": the H100 and H200 have been in mass production for years, Blackwell is ramping up a surge in volume, and Rubin announced full-scale mass production at this year's CES. The risks it faces are valuation, competition, and geopolitics, not whether the business model is viable.

On the technical side, it shows the market still recognizes this logic. The 50-day moving average is $206.52, and the 200-day moving average is $194.92. The stock price is about 15% above the 200-day line, with the moving averages in a bullish alignment. However, the 2.22 beta also warns: this is a stock that can amplify market sentiment by doubling.

Competitors are chasing single-card performance, but their moats are built in software and the power grid

AMD is the most direct challenger. The MI300X's 128GB HBM3 memory and MI350 have already approached or even caught up with the H200 in some inference benchmarks. The weak point has always been software—ROCm has been continuously improved for years, but official support for mainstream frameworks still lags behind CUDA by one or two versions.

The gap in scale is even harder to bridge with a single generation of products

AMD's data center business still generates annual revenue in the tens of billions of dollars, while NVIDIA alone generates $75.25 billion in data center revenue per quarter. This is not a conversation on the same scale.

Intel's Gaudi 3 targets mid-range training and inference, and oneAPI's cross-architecture approach is attractive to customers seeking open standards, but practical deployment cases are limited. Plus, its own process is still catching up with TSMC, so it poses no substantial threat in the short term.

The real concern comes from Broadcom's differentiated approach. It provides custom ASICs for Google's TPU and Meta's MTIA, combined with its own status as a switch chip, offering a complete interconnect solution—custom chips sacrifice versatility but deliver higher energy efficiency for specific loads.

Qualcomm has opened a new arena in edge inference. The Snapdragon X Elite's NPU is highly efficient on both PC and edge sides, but it hardly overlaps with data center GPU business.

So how wide is the moat? The gap lies in building three walls simultaneously on different materials: developers' time costs, full-stack integration tuning costs, and over twenty years of CoWoS advanced packaging capacity tied to TSMC. To replicate the opponent, they have to dig three different trenches at the same time.

The static ceiling has already been broken by its own revenue

Interestingly, the total market size allocated by many sell-side models for data center GPUs has long been caught up or even surpassed by NVIDIA's own revenue—when a company's annualized revenue exceeds $320 billion, the static TAM framework loses its relevance.

A more reliable algorithm is to add up the known numbers. Last quarter's actual revenue was $81.62 billion, and the company's median guidance for this quarter is $91 billion, meaning about $172.6 billion was locked in in the first half of this fiscal year alone; As long as there is no month-on-month contraction in the second half of the year, surpassing $350 billion for the year is almost a math problem.

Therefore, the ceiling is not on the demand side, but in the physical world. Third-party estimates show that the penetration rate of liquid-cooled AI chips will rise from 30% in 2025 to over 50% in 2026, and AI server shipments will still grow by 30% this year—but HBM memory supply, CoWoS packaging capacity, and grid and land approvals will determine how much growth can be realized.

NVIDIA has already begun to pry open these gates personally, which is why it has reached into power and real estate, which itself is both a widening moat and a new risk to its balance sheet.

Accelerated revenue, real cash, and a 20-year lease

The bulls' confidence isn't just narrative, but the fact that four groups have already fallen into the financial reports and announcements:

▲ Bull Case

Perfect pricing, single tenant, and $105 billion off-balance-sheet commitment

The evidence of the bears is also on the surface, with the latest one just being submitted to the SEC today:

▼ Bear Case

A quarterly report nine days later, and a lease that will wait until 2028

Short-term pacing is almost always driven by one main story, while the long-term story is staked on three other timelines:

•	August 26 US Stock Market After-Hours Trading: Last quarter's earnings report and new quarterly guidance release were the first to judge all the disagreements in this article
•	Second half of this year: The Rubin system was successively launched by AWS, Microsoft, and Google, and the six-chip complete machine solution began testing yield and delivery
•	Phased from 2028: The AI campus in Pike County, Ohio, will be commissioned in batches, with an initial 4.25 GW, verifying the off-balance-sheet guarantee realization path
•	Around 2027: HBM4e validation and the pace of Rubin Ultra mass production will directly determine whether the next generation gap can be maintained
•	Policy variables: The subsequent implementation of U.S. export controls and optical module import restrictions will affect the boundaries of accessible markets

From the crypto retreat to AI revaluation, valuations have been compressed and raised

Looking at the longer timeline, you'll find that the history of this stock is a cycle of expectations being constantly surpassed, valuations being repeatedly compressed and then risen. The 2019 acquisition of Mellanox laid the groundwork for the network, with P/E ratios exceeding 100 times from 2020 to 2021, only to be reverted to square one by the cryptocurrency retreat and rate hike cycles.

At the end of 2022, ChatGPT ignited demand for computing power, H100 supply was so tight that people had to queue, quarterly revenue multiplied several times in just over a year, and the stock price underwent a value revaluation accordingly. From 2024 to 2025, growth will shift from surprise to normal, with the market beginning to accept that "AI infrastructure is a long-term capital expenditure shift" rather than thematic hype.

This year, the script has shifted again: revenue growth has instead risen from 65% to 85%, but after the stock price rebounded from its yearly low to $225, it has remained stuck there—the market no longer applauds growth, but instead discounts on "how long growth can last." This silence itself is the most noteworthy signal to read right now.

⚠️ Risk Notice • $105 billion in off-balance-sheet guarantees are concentrated on a single tenant, creating significant impairment pressure in the event of a default • Valuations Cannot Slow Down: Free Cash Flow Yield Below 1.8%, Any Guidance Below Expectations Will Trigger Compression • TSMC's foundry and HBM supply are both centralized, and geopolitical or capacity events can directly interrupt shipments • Cloud vendors develop their own ASICs for continuous inference workloads, posing long-term market share risks of erosion • With a beta of 2.22 and the top weight in the index, passive capital flows amplify two-way fluctuations

🟡 Neutral Growth is still accelerating, but perfect pricing has drained the margin of safety.

💬 Discussion Growth is still ongoing, with limited room to maneuver; Real-time buying and selling positions can be found on the official WeChat account.

Data source

•	Source: NextPick real-time snapshot + company quarterly and annual reports + SEC announcement + cross-check by mainstream financial media.