NVIDIA has reached an unusual point in its investment story. The company is already one of the world's largest businesses, yet its latest quarter produced 106% revenue growth, Data Center revenue grew 117%, gross margins remained around 75%, and management guided the following quarter to approximately $108 billion of revenue.

At the same time, the stock has materially de-rated. Bloomberg recently reported that NVDA trades at less than 17× expected earnings over the next 12 months, near its lowest forward valuation in more than a decade. The central question for investors is therefore no longer whether NVIDIA dominates AI accelerators—it clearly does today—but whether the current AI infrastructure boom represents peak earnings or the beginning of a much larger cycle built around Rubin, inference, networking, sovereign AI and physical AI. My base case leans toward the latter, although hyperscaler spending, custom silicon, China and increasingly complex AI financing arrangements remain significant risks.

The NVIDIA thesis has evolved substantially. Three years ago, it could largely be summarized as AI needs GPUs, and NVIDIA dominates GPUs. Today, NVIDIA is increasingly selling an entire computing architecture encompassing accelerators, CPUs, NVLink, networking, software, AI factories and inference infrastructure.

That distinction is central to understanding the company's moat. NVIDIA has successfully moved customers from Hopper to Blackwell and Blackwell Ultra, while Vera Rubin is now ramping into production. The company is therefore becoming less dependent on simply producing the fastest individual accelerator and increasingly dependent on controlling the broader architecture required to build massive AI systems.

The long-term thesis rests on NVIDIA maintaining that position while global compute demand expands from AI model training into inference, agents, robotics, autonomous systems and enterprise AI. If that happens, NVIDIA's addressable market could continue expanding even as individual chips become more efficient and competing custom accelerators gain market share.

NVIDIA's latest reported quarter was fiscal Q2 2027, ended July 26. Revenue reached $96.22 billion, compared with $46.74 billion a year earlier, representing 106% year-over-year growth and 18% sequential growth. At its current quarterly revenue level, NVIDIA is approaching a roughly $400 billion annualized revenue run rate, which makes the continued triple-digit growth especially notable.

The business is also converting that growth into extraordinary profitability. Gross margins remained approximately 75%, while NVIDIA continues generating substantial operating leverage because operating expenses are growing far more slowly than revenue. This combination of enormous growth and exceptionally high margins is what separates NVIDIA from most semiconductor businesses.

Data Center generated approximately $89 billion in Q2, growing 117% year over year and 18% sequentially. It now represents roughly 92% of NVIDIA's total revenue, meaning NVIDIA should increasingly be viewed as an AI infrastructure company rather than a gaming-oriented semiconductor company.

Hyperscalers remain the most important source of demand. Amazon AWS, Microsoft Azure, Google Cloud, Meta, Oracle, CoreWeave and leading AI laboratories continue spending extraordinary amounts on AI infrastructure, and NVIDIA's hyperscale revenue more than doubled year over year during the quarter.

One particularly important recent example came from Amazon. AWS and NVIDIA announced in August that Amazon plans to deploy 2 million additional NVIDIA GPUs through fiscal 2029. This is notable because Amazon is simultaneously investing heavily in its proprietary Trainium accelerators. It suggests the future of AI infrastructure may not necessarily be NVIDIA or custom silicon; it may be NVIDIA and custom silicon as total compute demand expands.

Blackwell has become NVIDIA's current revenue engine, with Blackwell architecture now representing the vast majority of company revenue. The successful transition from Hopper into Blackwell and Blackwell Ultra removed one of the largest execution risks investors faced during the previous product cycle.

The next major transition is Vera Rubin, and this is arguably NVIDIA's most important catalyst over the coming year. Rubin isn't simply another GPU. The platform integrates the Rubin GPU, Vera CPU, NVLink, Spectrum-X networking, co-packaged optics and other system-level technologies into a broader AI computing architecture. NVIDIA says Rubin can provide approximately 10× greater agent throughput at scale than the previous Grace Blackwell generation.

Management expects Vera Rubin to represent approximately 20% of Data Center revenue in Q3, an unusually rapid contribution for a newly introduced architecture. NVIDIA also expects hyperscale growth to reaccelerate in Q4 and into fiscal 2028 as Rubin supply expands, with management offering a preliminary expectation for roughly 70% FY2028 revenue growth while supply remains constrained through at least the end of FY2028.

If those expectations hold, the current argument that NVIDIA is already approaching peak earnings could prove premature.

Training frontier AI models originally drove much of NVIDIA's explosive growth, but inference may eventually become the larger opportunity. Training happens periodically, while inference occurs every time a user asks an AI assistant a question, generates an image, runs an AI coding agent, performs AI search or interacts with an autonomous system.

That distinction becomes even more important with agentic AI. A traditional chatbot might generate a relatively small number of inference requests during a conversation. An autonomous agent performing research, coding, browsing or business workflows could make dozens or hundreds of model calls while completing a single task. The relevant demand metric therefore becomes less about how many AI models exist and more about how many tokens and computational steps are being generated globally.

This is one reason NVIDIA emphasizes Rubin's agentic-AI capabilities. If AI usage shifts from relatively simple chat interactions toward autonomous systems continuously performing work, inference compute consumption could become a second enormous growth engine independent of frontier-model training.

NVIDIA's greatest competitive advantage isn't simply GPU performance. It is CUDA and the broader software ecosystem surrounding NVIDIA hardware. Over many years, the company has built libraries, developer tools, optimized kernels, framework integrations, enterprise software and specialized AI tooling around its accelerators.

This creates substantial switching costs. A competitor cannot simply produce a processor that benchmarks slightly faster or costs slightly less. It needs competitive hardware, software, libraries, developer tools, support, networking, availability and system integration. Enterprises also need confidence that existing AI workloads can migrate without introducing major performance or development problems.

That ecosystem advantage is one of the biggest reasons NVIDIA has maintained such an extraordinary share of accelerated AI computing despite increasingly aggressive competition.

As AI clusters expand from thousands toward hundreds of thousands or eventually millions of accelerators, networking becomes increasingly important. The performance of an individual GPU matters less if processors cannot communicate efficiently across enormous clusters.

NVIDIA's networking portfolio—including NVLink, InfiniBand, Spectrum-X Ethernet and co-packaged optics—allows the company to capture more of the economics surrounding an AI data center. Instead of merely selling an accelerator, NVIDIA increasingly sells the architecture connecting accelerators together.

This creates an important expansion in wallet share. A customer can increasingly purchase the GPU, CPU, networking interface, switches, optical technology, software and support from NVIDIA. In effect, NVIDIA is moving from selling chips toward selling complete AI factories, strengthening both its competitive moat and the amount of revenue it can capture from every dollar of AI infrastructure spending.

NVIDIA's Q2 gross margin was approximately 75%, an extraordinary level for a semiconductor hardware company. Management expects Q3 gross margin around 74%, followed by a temporary decline toward approximately 71–72% in Q4, largely because of rising memory costs. Margins are then expected to recover toward roughly 72–73% during FY2028 as pricing adjustments take effect.

This is particularly relevant given the current memory shortage. The same HBM pricing environment producing extraordinary profitability for Micron is simultaneously increasing NVIDIA's input costs. That doesn't necessarily damage the long-term thesis, but it means investors should monitor whether NVIDIA can continue passing higher component costs through to customers without affecting demand.

The broader operating model remains exceptional. NVIDIA combines gross margins in the 70% range with operating expenses representing a relatively small portion of revenue, allowing enormous amounts of incremental revenue to convert into operating income.

NVIDIA generated approximately $96.6 billion of free cash flow in fiscal 2026, giving the company enormous financial flexibility. Unlike Amazon, Microsoft, Meta and Google—which must spend enormous amounts building the data centers that house AI infrastructure—NVIDIA operates a comparatively asset-light model and sells many of the highest-value components installed inside those facilities.

The company is increasingly returning that cash to shareholders. NVIDIA returned a record $26 billion during Q2, including approximately $20 billion of share repurchases and $6 billion of dividends. Management has indicated that it intends to return excess cash after strategic investment requirements and has targeted shareholder returns equivalent to at least 50% of free cash flow.

That represents an important evolution in the investment case. NVIDIA is no longer simply an exceptionally fast-growing technology company; it is becoming an enormous cash-generation and capital-return business simultaneously.

Gaming is no longer central to NVIDIA's valuation, but it remains a valuable secondary business. The Graphics segment generated approximately $7.9 billion during Q2, representing roughly 46% year-over-year growth. Blackwell-based GPUs also provide NVIDIA exposure to AI PCs, professional workstations, local inference and content creation.

These businesses matter because they diversify NVIDIA away from pure hyperscaler infrastructure demand, but the investment thesis today remains overwhelmingly dependent on Data Center. Investors should therefore avoid allowing strong gaming numbers to obscure deterioration in Data Center if that ever occurs.

Physical AI represents one of NVIDIA's most compelling long-duration opportunities. The company increasingly wants its computing architecture to become the standard platform for autonomous vehicles, humanoid robots, industrial robotics, warehouse automation, digital twins and automated factories.

The logic is similar to NVIDIA's original AI strategy. Today's AI largely exists inside computers and data centers, but future AI systems will increasingly interact with the physical world. Those machines will require substantial training infrastructure, simulation, inference hardware and software.

If NVIDIA establishes the same kind of ecosystem dominance in robotics that CUDA achieved in AI training, physical AI could eventually become another enormous addressable market. The opportunity remains much less financially material today than Data Center, but it provides valuable long-term optionality.

Governments increasingly view artificial intelligence as strategic infrastructure comparable to energy, telecommunications or defense technology. Countries want domestic compute capacity, sovereign data centers, locally trained models and greater control over sensitive data.

This creates a new category of NVIDIA customer beyond hyperscalers and private enterprises. Sovereign AI projects can potentially expand demand even if spending by traditional U.S. cloud providers eventually moderates. NVIDIA has continued expanding partnerships around sovereign and strategic AI infrastructure, reinforcing the idea that accelerated computing is becoming national infrastructure rather than simply another enterprise IT category.

One of NVIDIA's most significant recent strategic moves was its agreement to acquire Hugging Face for approximately $12.9 billion. Hugging Face is one of the most important platforms in open-source AI, providing models, datasets, developer tools and infrastructure used throughout the machine-learning ecosystem.

Strategically, the acquisition moves NVIDIA further up the software and developer stack. If NVIDIA owns critical infrastructure used by developers before workloads ever reach production, it can deepen relationships with the people ultimately deciding where AI workloads run.

There is also a legitimate risk. Hugging Face historically supports multiple hardware platforms, and developers could become concerned that NVIDIA ownership gradually favors NVIDIA hardware. Maintaining the platform's openness will therefore be important if NVIDIA wants to preserve its ecosystem value.

NVIDIA has also become increasingly involved in solving one of the industry's largest bottlenecks: financing AI infrastructure. The company partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish financing platforms targeting more than $500 billion of third-party capital for AI infrastructure.

Strategically, this makes sense. Demand for NVIDIA hardware can only become revenue if customers can finance the data centers, power infrastructure, networking and servers required to deploy it. Helping create financing mechanisms therefore potentially expands NVIDIA's addressable market.

However, this also introduces a new risk. NVIDIA increasingly invests in AI companies and infrastructure ecosystems that ultimately purchase NVIDIA hardware. Reuters recently highlighted concerns about increasingly complicated financing structures throughout the AI industry. Investors should therefore monitor how much end demand is supported by sustainable customer economics versus continuously expanding access to outside capital.

China remains one of NVIDIA's largest strategic risks. The company's latest filing showed that Data Center Hopper shipments to China represented less than 1% of Data Center revenue, while management excluded China Data Center compute revenue from forward guidance because of geopolitical uncertainty.

From a near-term earnings perspective, this is actually somewhat reassuring because current guidance doesn't depend heavily on a China recovery. Strategically, however, losing access to one of the world's largest technology markets remains meaningful.

The larger issue is that U.S. restrictions are accelerating China's effort to build an independent AI ecosystem. Huawei has accelerated its Ascend AI-chip roadmap and continues developing systems capable of connecting thousands of processors into increasingly large clusters. NVIDIA retains substantial advantages in software, performance and ecosystem maturity, but Huawei represents a genuine long-term competitor within China.

Over the long run, I view custom accelerators as a more important competitive risk than AMD alone. NVIDIA's largest customers increasingly design proprietary chips: Google has TPU, Amazon has Trainium, Microsoft has Maia and Meta has MTIA, while Broadcom and others help hyperscalers develop additional custom ASICs.

The motivation is obvious. At enormous scale, hyperscalers want lower costs, better energy efficiency, chips optimized for specific workloads and less dependence on NVIDIA.

However, Amazon's recent commitment to deploy 2 million additional NVIDIA GPUs despite simultaneously expanding Trainium demonstrates why custom silicon does not automatically break the thesis. The likely future is a mixed architecture where NVIDIA dominates flexible, cutting-edge accelerated computing while proprietary ASICs capture workloads that can be highly optimized.

The important question is therefore not whether custom silicon gains share—it almost certainly will—but whether total AI compute demand grows quickly enough for NVIDIA revenue to continue expanding despite that share loss.

AI models and inference systems will almost certainly become more efficient over time. On the surface, that appears bearish for NVIDIA because customers may require fewer GPUs to perform the same amount of computation.

But falling compute costs can also dramatically increase usage. If the cost of inference declines 90%, businesses may respond by running far more AI agents, integrating AI into more workflows, putting AI on more devices and deploying substantially more autonomous systems.

This is essentially a Jevons-paradox argument. Greater efficiency can increase total consumption when lower costs unlock previously uneconomic applications. For NVIDIA, the relevant equation is therefore not simply compute cost declines; it is whether growth in AI usage exceeds the decline in cost per unit of intelligence.

The market's current skepticism reflects several legitimate concerns: enormous hyperscaler capex commitments, increasingly complex AI financing, rapidly improving Chinese models, custom accelerators and uncertainty around whether end customers can generate adequate returns from AI infrastructure.

Those concerns don't invalidate NVIDIA's current results. Demand remains exceptionally strong and supply remains constrained. But they explain why investors are willing to assign NVIDIA a substantially lower forward multiple despite triple-digit revenue growth.

The market is effectively asking whether NVIDIA's current earnings are sustainable or whether hyperscalers are temporarily overbuilding infrastructure ahead of economically justified demand. That is the most important debate surrounding the stock.

NVIDIA recently traded around $220, while recent Wall Street price targets have ranged widely. Recent targets include Piper Sandler at $300, JPMorgan at $320, RBC around $330, Bernstein around $400 and Evercore substantially higher. Recent consensus compilations have generally fallen in the low-to-mid $300 range.

More interesting than analyst targets is NVIDIA's multiple compression. Bloomberg recently reported that NVIDIA trades below 17× next-12-month expected earnings, close to its lowest forward valuation in more than a decade.

That valuation is striking against the current fundamentals: revenue growth of 106%, Data Center growth of 117%, roughly 75% gross margins, Q3 revenue guidance near $108 billion and management's preliminary expectation for approximately 70% FY2028 revenue growth.

The market clearly isn't questioning whether NVIDIA is currently performing well. It is questioning how long those economics can persist.

The most important near-term catalyst is the Rubin ramp. A successful transition that quickly turns Rubin into a substantial portion of Data Center revenue would demonstrate that NVIDIA can continue moving customers through rapid architecture cycles without creating demand pauses.

Beyond Rubin, the primary catalysts are continued hyperscaler capex growth, accelerating inference workloads, agentic AI adoption, networking growth, sovereign AI investment, expansion into physical AI and increased monetization of NVIDIA's software ecosystem. If several of those opportunities mature simultaneously, NVIDIA could sustain unusually high growth even as the traditional training market becomes more mature.

Free cash flow and shareholder returns provide another catalyst. Continued earnings growth combined with aggressive buybacks could increasingly support EPS even as NVIDIA becomes a much larger company.

The largest risk is an AI capex slowdown. If Amazon, Microsoft, Meta, Google, Oracle and major AI laboratories determine that AI infrastructure isn't producing adequate returns, NVIDIA would likely feel the impact quickly.

Custom ASICs represent the next major structural risk, particularly as NVIDIA's largest customers seek lower-cost alternatives. China restrictions reduce NVIDIA's addressable market while simultaneously accelerating Huawei's development. Gross margins face pressure from rising HBM and other component costs, and customer concentration means a relatively small group of hyperscalers accounts for an enormous amount of industry spending.

Investors should also monitor the increasingly complicated financing arrangements surrounding AI infrastructure. Demand supported by profitable end-user workloads is much healthier than demand supported primarily by cheap or continually expanding access to capital.

Finally, NVIDIA must execute an increasingly aggressive product cadence. The company has successfully transitioned from Hopper to Blackwell, but maintaining annual architecture transitions without manufacturing, software or supply-chain disruptions becomes more difficult as systems become increasingly complex.

This is the portion I would update after every NVIDIA earnings report:

I would become substantially less constructive if Data Center growth falls below roughly 25–30%, hyperscalers begin cutting AI capex, Rubin experiences material delays, gross margins remain sustainably below 70%, NVIDIA begins losing major workloads to custom silicon faster than overall AI demand grows, inventories rise significantly faster than revenue, or cloud providers report declining AI infrastructure utilization.

The single most important signal would be evidence that NVIDIA's customers aren't generating adequate economic returns from AI infrastructure. Ultimately, NVIDIA's customers must earn enough money from AI to justify continuously purchasing increasingly expensive computing systems. If customer economics deteriorate, almost every other part of the bull thesis becomes less important.

Rather than assuming today's triple-digit growth persists indefinitely, I would value NVIDIA around normalized forward earnings and the multiple investors are likely to assign those earnings as growth eventually moderates.

The bear case assumes AI infrastructure spending peaks, hyperscaler capex slows, custom ASIC adoption accelerates, China remains largely inaccessible and gross margins normalize below 70%. NVIDIA would remain an exceptional company, but earnings estimates would fall while investors begin valuing the business more like a mature semiconductor leader.

My base case assumes Rubin ramps successfully, Data Center remains extremely strong, inference becomes a much larger contributor to total compute demand, hyperscaler capex stays elevated and custom silicon gains share without breaking NVIDIA's CUDA ecosystem. Gross margins stabilize around the low-to-mid 70% range after temporary memory-cost pressure.

Under those assumptions, I believe NVIDIA can justify roughly a mid-20s multiple on normalized forward earnings, producing a valuation around $320 over the next 12 months.

The bull scenario requires NVIDIA to remain the overwhelmingly dominant AI infrastructure platform while Rubin significantly expands revenue, inference usage explodes, AI agents increase compute consumption, networking becomes a substantially larger business, physical AI begins becoming financially meaningful and NVIDIA's software ecosystem continues expanding.

Under that scenario, NVIDIA increasingly deserves to be valued as an AI infrastructure platform rather than a traditional semiconductor manufacturer.

The $320 target assumes NVIDIA successfully transitions from Blackwell Ultra into Vera Rubin, Data Center growth remains exceptionally strong, hyperscaler AI spending stays elevated, inference becomes a larger contributor to global compute demand and gross margins stabilize in the low-to-mid 70% range after near-term memory-cost pressure.

What makes the current setup particularly compelling is the combination of extraordinary fundamentals and substantial multiple compression. NVIDIA's latest quarter delivered 106% revenue growth and 117% Data Center growth, yet the stock's forward earnings multiple has fallen below 17× according to Bloomberg.

The largest threat isn't simply AMD releasing a competitive accelerator. It's that NVIDIA's customers eventually stop increasing AI infrastructure spending at anything resembling today's rate. Custom silicon, Huawei, China restrictions, memory costs and financing conditions all matter, but hyperscaler return on invested capital remains the variable I would watch most closely.

For now, the fundamental evidence remains supportive. Rubin is ramping, AWS has committed to millions of additional NVIDIA GPUs, supply remains constrained, and management continues to expect exceptionally strong growth into fiscal 2028. That supports my current BUY framework while leaving plenty of reasons to monitor the thesis closely as the AI infrastructure cycle matures.

Newsletter Verdict: BUY — $320 12-month target.

This rating and valuation are an equity-research framework based on the assumptions above, not individualized investment advice.