Signals Inbox·July 28, 2026·AI Chips

Is Nvidia getting less or more powerful?

Nvidia is getting more powerful even as its dominance becomes more contested: it is selling far more infrastructure, capturing more of each data center and setting wider technical standards, while its biggest customers race to reduce their dependence on it.

Q1What does “more powerful” actually mean for Nvidia?

Nvidia is more powerful when it can make customers spend more, adopt its technical standards and build data centers around its roadmap, even when its share of individual AI chips falls.

That definition is more useful than asking whether Nvidia still controls almost every AI accelerator sale. Google, Amazon, Meta and Microsoft are all building their own chips, while AMD now has credible large-scale customers. Nvidia will almost certainly face a lower market share as these alternatives spread.

Yet market share tells only part of the story. Nvidia can lose ten percentage points of a rapidly expanding market and still sell many more systems, earn more profit and gain influence over how the industry operates. It can also capture more money from each installation by selling networking equipment, CPUs, switches, storage processors, software and complete racks alongside its GPUs.

Nvidia’s financial and technical power has grown quickly. Its freedom has narrowed in China, its largest customers are funding alternatives, and its supply chain remains concentrated in Asia. Power is rising, but so is the resistance around it.

Q2Is Nvidia getting economically stronger right now?

Nvidia is becoming much stronger economically, and the latest results show no sign that its growth has settled into a normal semiconductor pace.

Quarterly revenue reached $81.6 billion, according to Nvidia’s latest financial results. That was 85% higher than one year earlier and 20% higher than the previous quarter. Data-center revenue reached $75.2 billion, up 92% in twelve months.

The scale of the change is easier to see across several years. Nvidia generated $60.9 billion during its entire fiscal 2024. Two years later, annual revenue had reached $215.9 billion. The latest single quarter was already one-third larger than that earlier full year.

Profitability has risen alongside revenue. Nvidia produced $53.5 billion in operating income during the latest quarter, giving it an operating margin of roughly 66%. Its reported net income was even higher, although a large unrealized gain on Nvidia’s investment portfolio inflated that figure. Operating income gives us the cleaner picture of the underlying business.

Hardware companies rarely expand this quickly while preserving such margins. Product transitions, manufacturing bottlenecks and large customer negotiations normally pull profitability down. Nvidia has absorbed those pressures while increasing both volume and the amount it earns from each AI system.

The numbers are blunt: Nvidia’s economic power is growing at a speed its competitors have yet to match.

Nvidia revenue, operating income and research spending

Q3Is Nvidia still the default AI infrastructure company?

Nvidia is still the default choice for large, general-purpose AI systems, although customers now have credible reasons to use other chips for selected workloads.

A default platform does not need to win every purchase. It needs to be the option customers can choose with the least technical and commercial risk.

Nvidia still has that advantage. Its hardware is available through Amazon Web Services, Microsoft Azure, Google Cloud, Oracle and specialist AI clouds. Frontier laboratories including OpenAI, Anthropic and xAI build major parts of their computing infrastructure around Nvidia systems. Enterprises can buy similar technology from Dell, HPE, Lenovo, Supermicro and a long list of server manufacturers.

That reach creates a convenience rivals struggle to match. A company can train on Nvidia hardware in one cloud, move to another provider, hire engineers who already know CUDA and find established tools for monitoring, networking and inference. Switching to a custom accelerator can reduce costs, but it usually narrows the company’s choice of clouds and requires more engineering work.

The newest deployments suggest that this default status remains intact. Microsoft’s planned Fairwater systems are expected to use hundreds of thousands of Vera Rubin Superchips. Meta announced a long-term Nvidia infrastructure partnership even while expanding its own chips and placing a major order with AMD. Vera Rubin racks are also being prepared across more than 350 manufacturing sites in 30 countries.

Nvidia’s competitors now offer real alternatives. None currently offers the same combination of hardware availability, software maturity, cloud support and trained developers.

Q4Can Nvidia keep charging premium prices?

Nvidia can still charge premium prices because buyers care more about how quickly a working cluster produces useful results than about the price of one accelerator.

Its latest gross margin was 74.9%, and its guidance kept the expected margin around 75%. That level has survived enormous growth, the shift from Hopper to Blackwell and the rising complexity of complete rack-scale systems.

A customer comparing chips may find a lower unit price elsewhere. The final bill also includes engineering time, networking, power, cooling, utilization, maintenance and the cost of waiting for a new system to work reliably.

Nvidia has become particularly good at charging for time saved. It sells a pre-integrated combination of processors, interconnects, network switches, software libraries and reference designs. The GPU price is only one line in a much larger decision.

Meta, Google, Amazon and Microsoft have enough scale to negotiate aggressively, while AMD’s Helios systems and internal chips such as TPU and Trainium create more pressure on individual GPU prices. Nvidia’s wider system pricing looks harder to break.

Q5Are Nvidia’s biggest customers gaining the upper hand?

Nvidia’s largest customers have more leverage than before, but they still need Nvidia badly enough to prevent that leverage from turning into control.

The hyperscalers can influence orders, delivery schedules and system design simply because they buy at extraordinary scale. They also own the cloud platforms through which many businesses access Nvidia hardware.

Their strongest weapon is internal chip development. Google has TPU, Amazon has Trainium, Microsoft has Maia and Meta has MTIA. These projects give each company a fallback option when Nvidia’s price, supply or roadmap does not fit a particular workload.

The spending behind them is substantial. Alphabet currently expects capital expenditure of $180 billion to $190 billion for the year and says investment will rise significantly again next year. Meta is developing four new MTIA generations in two years. Amazon keeps broadening support for Trainium across its cloud services.

And yet these companies continue signing large Nvidia agreements. Meta expanded its Nvidia partnership while committing to AMD and building MTIA. AWS promotes Trainium and Nvidia GPUs side by side. Google Cloud is increasing TPU availability while also preparing to run Rubin systems.

They are reducing dependence without giving up access to the platform their own customers still want.

Q6Are custom AI chips replacing Nvidia?

Custom chips are taking specific workloads away from Nvidia, especially high-volume inference, but they are far from replacing it as the shared platform of the AI industry.

Meta offers one of the clearest examples. The company says it already runs hundreds of thousands of MTIA chips across advertising and organic-content inference. MTIA 300 is in production, while MTIA 400, 450 and 500 are intended to handle recommendations, training and generative AI inference.

That is real deployment rather than an experimental chip program. Meta can use MTIA efficiently because it controls the models, applications, data centers and software. Even a modest saving becomes valuable when repeated across billions of recommendations and advertisements.

Google is moving in a similar direction. It has used TPUs internally and through Google Cloud for years. It now plans to deliver TPU hardware into selected customers’ own data centers, with most of the expected revenue from those agreements arriving later. That turns TPU into a direct infrastructure product rather than a feature available only inside Google’s cloud.

Amazon’s approach follows the same logic. Trainium and Inferentia target customers willing to use AWS software in exchange for lower training or inference costs. Amazon has reported cases where its own services achieved materially lower inference costs on those chips.

Custom chips are strongest when the workload is predictable, enormous and controlled by one platform owner. Nvidia remains stronger when customers want flexibility across models, clouds and use cases.

This will probably cost Nvidia market share. It may also expand the overall market by making some AI workloads cheaper to run.

Q7Has AMD finally become a serious Nvidia rival?

AMD is now a serious Nvidia rival, although most of its biggest AI commitments still need to turn into operating systems and reported revenue.

The latest change is AMD’s move from selling individual accelerators to offering a complete rack-scale platform. Helios combines 72 MI455X GPUs, 18 EPYC CPUs, Pensando networking and ROCm software. AMD says the system is already in production and can deliver up to 30% more inference tokens per dollar than the leading competing system.

That 30% figure comes from AMD’s own testing, so independent production results still matter. The surrounding commercial progress is harder to dismiss.

OpenAI has agreed to deploy up to six gigawatts of AMD GPUs, beginning with an initial one-gigawatt installation. Meta has a separate agreement covering as much as six gigawatts. Anthropic has announced plans for up to two gigawatts. Microsoft, Oracle and several specialist clouds are also preparing Helios deployments.

These commitments show that sophisticated customers now believe AMD can become a second large-scale supplier. They want lower prices, more supply and less dependence on one roadmap.

The current financial gap remains huge. AMD’s latest data-center revenue was $5.8 billion, and that figure includes server CPUs alongside AI accelerators. Nvidia’s comparable data-center business was roughly 13 times larger.

AMD has reached the stage where it can change Nvidia’s behavior. It can win major projects, give buyers negotiating leverage and establish more open standards. Catching Nvidia’s installed base, software ecosystem and revenue scale will take much longer.

Nvidia and AMD in large-scale AI infrastructure

Q8Is CUDA still Nvidia’s strongest moat?

CUDA remains Nvidia’s strongest moat, although its value increasingly comes from the whole software stack built around it rather than the programming layer alone.

Nvidia says its ecosystem now includes more than six million developers and nearly 6,000 accelerated applications. Those headline numbers are only a rough measure of the advantage. The deeper value lies in years of libraries, tutorials, debugging tools, optimized kernels and internal company workflows.

A model may technically run on AMD or a TPU while still performing worse than it did on Nvidia. Engineers then need to identify slow operations, rewrite code, test numerical differences and rebuild deployment tools. Large AI laboratories can afford that work. A normal enterprise often prefers to pay more for Nvidia and avoid it.

Competitors have made portability easier. PyTorch, Triton, vLLM and open model formats reduce the amount of code tied directly to one chip. Meta deliberately built MTIA around these industry standards. Google is improving native PyTorch support for TPU, while AMD continues investing in ROCm.

Nvidia has answered by extending CUDA into more parts of the system. Its advantage now includes distributed training libraries, model-serving software, networking tools, inference engines, security and cluster management. A competitor may support the application while still lacking equally mature tools underneath it.

CUDA’s grip will weaken gradually as the largest customers demand portability. For the broader market, choosing Nvidia is still the easiest way to avoid spending months solving infrastructure problems that competitors have not fully standardized.

Q9Will the move toward AI inference weaken Nvidia?

Inference will weaken Nvidia’s share of some AI workloads, but the enormous growth in daily AI usage should still make Nvidia’s inference business much larger.

Training a frontier model is a concentrated task. It requires a huge cluster for a limited period and rewards maximum flexibility. Inference runs continuously after the model launches, sometimes serving billions of requests. Costs, latency and power consumption become much more important.

That environment suits custom chips. Meta can optimize MTIA for recommendations and generative AI. Amazon can tune Trainium and Inferentia for its cloud software. Google can run Gemini and Search on hardware designed around its own workloads.

Inference is also highly fragmented. A coding assistant, advertising model, video generator and voice agent need different combinations of memory, speed and precision. There may be no single chip that wins all of it.

Nvidia is responding by designing whole systems around the cost of generating tokens. Vera Rubin combines several processor, networking and storage designs, while Nvidia’s newer software manages context, routing and distributed inference. The company claims large gains in agent throughput and token economics compared with Blackwell, though real results will vary by model.

Nvidia’s share should come down from its unusually high level. Daily inference demand can still make the business much larger.

Q10Is Nvidia becoming an entire data-center company and shaping how AI facilities are built?

Nvidia is now much closer to an AI data-center company than a traditional GPU supplier, and its systems increasingly shape the technical design of the buildings around them.

Its latest quarter included $60.4 billion of data-center compute revenue and $14.8 billion from networking. Networking grew 199% in one year and represented almost one-fifth of data-center sales.

The reason is simple. Tens of thousands of powerful GPUs are useless when they cannot communicate quickly enough. As clusters become larger, the connections between chips determine how much time and electricity are wasted.

Nvidia now sells most of the important pieces around that problem. Vera Rubin includes GPUs, Vera CPUs, NVLink switches, Spectrum Ethernet equipment, BlueField data-processing units and storage infrastructure. Nvidia’s software then coordinates work across those components.

Its rack designs affect buildings before the chips arrive. They determine power density, cooling systems, network layouts, cable requirements and the space needed around each installation. Nvidia’s DSX reference design extends that influence into the physical facility, infrastructure software, simulation and operations.

This expansion raises the value of each deployment and the barrier for rivals. AMD must compete with an integrated rack, while cloud providers must decide how much of their architecture they are comfortable sourcing from one company.

Nvidia still depends on decisions made elsewhere. Microsoft, Amazon, Google, Meta, Oracle and CoreWeave control the capital budgets and customer relationships. Utilities decide whether enough power is available, and governments influence permits, grid connections and exports.

Nvidia currently writes much of the technical blueprint. It does not own the land, electricity or cloud service wrapped around that blueprint.

Q11Can Nvidia’s yearly product roadmap keep competitors behind?

Nvidia’s yearly roadmap is keeping competitors under pressure, but the growing complexity of each launch makes a serious execution mistake more likely.

The company moved from Hopper to Blackwell and then to Vera Rubin in a remarkably short period. Rubin combines several new processor, networking and storage products that must work together as one system.

This pace makes it difficult for competitors to target Nvidia. By the time a rival catches the performance of one GPU, customers may already be planning buildings and software around the next complete rack.

Recent production updates suggest that Rubin remains on schedule. Nvidia says manufacturing is ramping through hundreds of partners, with initial shipments expected to begin later this year. Systems are already running at several cloud partners as the production network expands.

Speed has a cost. Nvidia’s latest quarterly filing says engineering development material expenses increased 204% from one year earlier. The company is also preparing $32.4 billion of future data-center lease obligations, mainly to support its own research and development.

Supply commitments create another risk. Nvidia ended its previous fiscal year with tens of billions of dollars tied to inventory, manufacturing capacity and future purchases. Those agreements help it reserve scarce components before rivals can obtain them. They become painful when demand changes or a new architecture arrives late.

Blackwell already showed how one design issue can disrupt supply planning and create inventory charges. Rubin combines even more components.

The annual cadence is a powerful weapon while Nvidia executes well. One badly delayed generation would give customers a rare opportunity to shift budgets toward AMD or their own chips.

Q12Is Nvidia too dependent on TSMC and HBM?

Nvidia remains heavily dependent on TSMC, advanced packaging and high-bandwidth memory, and no realistic short-term plan can remove that dependence.

Nvidia designs its chips but relies on outside companies to manufacture, package and equip them with memory. TSMC sits at the center of that chain. SK Hynix, Samsung and Micron supply high-bandwidth memory, while Asian manufacturing partners assemble the final systems.

These components cannot be swapped quickly. Moving a complex processor to another foundry would require redesign, testing and a long production ramp. A shortage of memory, packaging capacity, optical components or substrates can delay a complete rack even when the GPU itself is ready.

TSMC’s latest results show how valuable this bottleneck has become. Quarterly revenue reached $40.2 billion, up nearly 34% in US dollar terms, with a gross margin of 67.7%. Advanced processes of 7 nanometers or smaller generated 77% of wafer revenue.

Those margins show that TSMC has bargaining power of its own. Nvidia is one of its largest growth engines, but TSMC serves many leading chip designers and controls manufacturing expertise that Nvidia cannot recreate quickly.

Nvidia has responded by reserving capacity far in advance, qualifying more suppliers and spreading final system manufacturing across hundreds of factories. Samsung’s progress in newer HBM generations also gives Nvidia more options alongside SK Hynix.

The dependence is manageable during normal operations. A disruption involving Taiwan, South Korea or advanced packaging would reveal how little direct control Nvidia has over the physical production of its most important products.

Q13Has losing China made Nvidia less powerful?

Nvidia is clearly less powerful in China, even though demand elsewhere has more than covered the immediate revenue loss.

The company says it is effectively blocked from competing in China’s data-center computing market under the current combination of US export controls and Chinese policy.

The damage is measurable. Earlier restrictions forced Nvidia to take a $4.5 billion charge related to H20 inventory and purchase commitments. In the latest quarter, it shipped no Hopper data-center products to China, compared with $4.6 billion one year earlier.

The United States has granted licenses for small quantities of H200 systems to selected Chinese buyers. Nvidia says it has generated no revenue from that program so far and remains unsure whether China will allow the imports. Any permitted H200 would also pass through a US inspection process and face a 25% tariff.

Nvidia’s current revenue forecast assumes no data-center compute sales to China. The company can still grow quickly because buyers in the United States, the Middle East, Europe and Asia are absorbing available supply.

The longer-term loss is more serious. Chinese companies now have a stronger reason to improve domestic chips, software and networking. Developers who would previously have built around CUDA are learning to make models work on Huawei and other local platforms.

Every year Nvidia remains absent gives that ecosystem more users, engineers and production experience. Nvidia is becoming more powerful across much of the world while one of the largest potential AI markets moves further beyond its reach.

Q14Could the AI spending boom turn against Nvidia?

A sharp slowdown in AI infrastructure spending would hurt Nvidia badly, but current evidence still looks more like a race for scarce computing capacity than a market preparing to collapse.

The spending plans are enormous. Alphabet expects capital expenditure of $180 billion to $190 billion this year and says next year will be significantly higher. Meta, Microsoft and Amazon are also committing tens of billions of dollars to chips, buildings, networking and power.

Some of that money goes to custom silicon, yet Nvidia remains one of the largest beneficiaries. The company’s forecast calls for approximately $91 billion of revenue in the current quarter despite including no Chinese data-center compute sales.

Customers are beginning to feel the financial weight. Depreciation, electricity and data-center operating expenses are rising. Amazon’s investment has put pressure on free cash flow. Alphabet has warned that infrastructure costs will keep increasing. Microsoft has said AI infrastructure is affecting cloud margins.

Those pressures can eventually force buyers to demand better utilization and slower expansion. A project with weak usage becomes harder to justify once the first excitement around AI capacity fades.

Demand is producing revenue now, though. Google Cloud’s generative AI products are growing rapidly, Microsoft’s AI business has reached a large annual run rate, and the major cloud providers still describe themselves as capacity constrained.

The dangerous combination for Nvidia would be weaker AI revenue, overbuilt data centers and a rapid shift toward cheaper internal chips. We do not currently see all three together.

A correction is plausible. A collapse is not supported by the behavior of the largest customers.

Q15Are regulators becoming a real limit on Nvidia?

Regulators are starting to restrict Nvidia’s freedom, although trade policy has inflicted far more damage than competition investigations so far.

Nvidia’s position has attracted attention in the United States, European Union, United Kingdom, China, South Korea, Japan and France. Its filings describe requests for information covering GPU sales, supply allocation, partnerships, investments and agreements with AI model developers.

The concerns are understandable. Nvidia supplies the dominant computing platform, invests in companies that purchase its hardware and works closely with cloud providers that resell it. Its control is also spreading into networking and complete systems, making the boundaries of the market harder to define.

No major antitrust remedy has yet changed Nvidia’s core business. Investigations can consume management time and make certain agreements harder to structure, but customers are still signing large contracts and Nvidia is still expanding into adjacent products.

Export controls have had an immediate effect. They removed most of Nvidia’s Chinese data-center opportunity, created inventory charges and encouraged Chinese customers to build around local hardware. Few actions by a rival company have damaged Nvidia that quickly.

The next area to watch is Nvidia’s growing investment network. The company invested $18.6 billion in private companies and infrastructure funds during the latest quarter. Some recipients may later buy or use Nvidia systems, which can blur the line between independent demand and demand supported by Nvidia’s own capital.

Regulation has become a meaningful constraint. It has not yet weakened Nvidia’s commercial power outside China.

Q16Is Nvidia’s power spreading beyond the hyperscalers?

Nvidia’s power is spreading beyond the largest cloud companies, and its latest revenue breakdown shows that the expansion is already financially important.

Nvidia recently divided its data-center business into two groups. Hyperscale customers produced $37.9 billion in the latest quarter. AI clouds, industrial companies, enterprises and sovereign buyers produced $37.4 billion.

The two sides were almost equal. Hyperscale revenue grew faster, at 115% year over year, while the broader group still grew 74%.

This helps answer an important concern about Nvidia’s dependence on a handful of technology giants. Amazon, Microsoft, Google and Meta remain central buyers, but half of data-center revenue now comes from a wider set of customers.

The recent projects also look different from the early AI boom. Japan is preparing a national physical-AI system with 27,500 Rubin GPUs. Pharmaceutical companies are building specialized AI factories for drug research. Governments, telecom operators, industrial groups and robotics companies are adopting Nvidia’s software and hardware together.

These deployments will not match hyperscaler purchases individually. Their combined value lies in extending Nvidia’s standards into more industries. A manufacturer using Nvidia simulation, robotics and data-center systems becomes harder to win away than a cloud company buying accelerators through a competitive tender.

Nvidia still relies on a limited number of direct distributors and system builders to deliver many of these sales. The end demand, however, is becoming broader.

That diversification makes Nvidia less exposed to the spending decision of any single cloud company and gives CUDA, networking and Rubin a larger installed base.

Q17So, is Nvidia getting less or more powerful?

Nvidia is getting more powerful overall, although its customers, competitors and governments are pushing back much harder.

The financial verdict is decisive. Quarterly revenue is growing by 85%, data-center sales have nearly doubled in one year, and gross margin remains close to 75%. Nvidia is capturing revenue from complete AI systems rather than relying only on individual GPUs.

Its technical reach is also wider. CUDA remains the easiest shared platform for many developers. Rubin extends Nvidia into CPUs, networking, storage and data-center design. Demand is spreading from hyperscalers into AI clouds, governments and industrial companies.

The forces working against Nvidia are stronger than they used to be. AMD has real gigawatt-scale commitments. Meta runs hundreds of thousands of its own accelerators. Google is preparing to sell TPU hardware for customer data centers. Amazon is expanding Trainium. China has largely closed its data-center market to Nvidia.

These developments should reduce Nvidia’s accelerator share and give large buyers more negotiating room. They have yet to produce lower Nvidia revenue, weaker margins or a broad customer retreat.

Market share and power can move differently. Nvidia may control a smaller percentage of future AI chips while earning more money, supplying more of each data center and setting more of the industry’s technical standards.

The claim that Nvidia is becoming less powerful is premature. Nvidia is currently more powerful economically, technically and commercially. What it has lost is the period when serious resistance barely existed.

Where Nvidia is gaining and losing power