Signals Inbox·July 28, 2026·AI Infrastructure

Can CoreWeave beat Amazon and Microsoft?

No single company is likely to kill Nvidia. The credible threat is a coalition in which Google and Amazon move major workloads onto their own chips, AMD becomes the usable second supplier, Broadcom equips the custom-silicon push, and open software makes switching less painful.

Q1What would it actually mean to kill Nvidia?

Killing Nvidia would mean making its platform optional for serious AI work.

The company would not need to disappear. Nvidia could remain huge in gaming, robotics, automotive systems and scientific computing while losing the position that currently lets it set the pace and capture an exceptional share of AI infrastructure spending.

We would consider Nvidia strategically defeated if three changes happened together. Large model developers could train and serve their most important systems without depending on Nvidia. Engineers could move workloads between accelerators without months of rewriting and tuning. Customers would also have enough credible options to push Nvidia's margins much closer to normal semiconductor levels.

A few lost points of market share would barely change the story. Nvidia currently sells the accelerator, the networking, much of the software and increasingly the rack surrounding the accelerator. A serious challenger therefore has to weaken the wider platform or make several parts of it unnecessary.

Here, “kill” means turning Nvidia from the default AI computing platform into one powerful supplier among several.

Q2Why does the Nvidia killer question feel real now?

The Nvidia killer question feels real now because alternatives are moving from technical trials into production infrastructure.

Several recent developments point in the same direction. Microsoft is expanding Azure around AMD's Helios platform and MI455X accelerators for production inference. Google's seventh-generation Ironwood TPU can operate in pods containing as many as 9,216 chips. Anthropic has also deepened its relationship with Amazon and reserved several gigawatts of additional AWS capacity.

Nvidia is not losing today. Its latest reported quarter was its strongest ever, and management guided toward even higher revenue in the following one. What has changed is the way its largest buyers operate. Google, Amazon, Microsoft, Meta and OpenAI increasingly divide their computing needs between Nvidia, internal silicon, AMD and more specialized accelerators.

A few years ago, this diversification mainly looked like a bargaining tactic. Now it is becoming an infrastructure strategy. The largest AI buyers have enough internal expertise to decide, workload by workload, which hardware deserves the business.

Q3How far ahead is Nvidia today?

Nvidia is still miles ahead today.

Its latest quarter produced $75.2 billion in data-center revenue, including $60.4 billion from compute and $14.8 billion from networking. AMD reported $5.8 billion for its entire data-center segment, which also contains server processors and products beyond AI accelerators. Broadcom generated $10.8 billion from AI semiconductors, including custom accelerators and AI networking.

The gap becomes obvious when we put those numbers together. Nvidia's quarterly data-center revenue was about thirteen times AMD's. Nvidia also spent $6.3 billion on research and development during the quarter, more than AMD generated from its whole data-center division. That difference lets Nvidia fund new GPUs, CPUs, switches, interconnects, software and complete rack systems at the same time.

Its pace is just as striking. Nvidia's data-center revenue rose 92% from the previous year, while management forecast roughly $91 billion of total revenue for the following quarter. Every challenger is chasing a leader whose business is still accelerating.

Latest reported quarterly measures

Q4Can AMD actually catch Nvidia?

AMD can become a powerful number two, but it cannot kill Nvidia by itself.

AMD has the clearest role among Nvidia's direct rivals because it sells a general-purpose accelerator that clouds, enterprises and model developers can buy. Customers do not need to design their own chips or tie all their workloads to one cloud. ROCm has also matured into a usable software platform, PyTorch supports it directly, and AMD's HIP tools make much CUDA-style code easier to move.

The commercial commitments are now large enough to take seriously. OpenAI and Meta have each agreed to deploy up to 6 gigawatts of AMD accelerators over several years. Oracle plans a cluster containing 50,000 MI450 chips. Microsoft is also adding AMD Helios systems and MI455X virtual machines to Azure for production AI inference.

Those agreements have changed AMD's position. A few years ago, customers often treated Instinct as a backup option or a way to negotiate with Nvidia. AMD is now part of long-term infrastructure plans, sometimes through custom versions of its chips and systems designed around a customer's own workloads.

The remaining gap is brutal. AMD's $5.8 billion in quarterly data-center revenue includes EPYC CPUs as well as Instinct accelerators, while Nvidia generated more than ten times that amount from data-center compute alone. AMD must also prove that its large commitments become installed capacity, reliable utilization and recurring revenue. Multiyear projects can still slip because of software problems, manufacturing constraints or changing customer demand.

AMD is most likely to hurt Nvidia first on price. It can take enough business to stop Nvidia charging every customer as though no alternative exists. Breaking the broader platform will also require progress in open software, open networking and custom chips.

Q5Is Google the strongest single threat to Nvidia?

Google is currently the company most capable of challenging Nvidia on its own.

It already controls every layer required to run a major AI business without Nvidia: frontier models, a global cloud, custom chips, compilers, networking, data centers and enormous internal demand. Google has also kept improving the TPU for almost a decade instead of launching one defensive chip and quietly abandoning the effort.

Ironwood shows how far that program has progressed. A full pod connects 9,216 chips, provides 1.77 petabytes of shared high-bandwidth memory and reaches 42.5 exaflops. Google says each chip offers roughly four times the performance of the previous Trillium generation, with particularly large economic gains on some low-latency mixture-of-experts inference workloads.

The deployment history is more important than the specifications. Google has trained and served several generations of Gemini on its own infrastructure, then opened TPUs to outside customers through Google Cloud. Anthropic has expanded its use of Google TPUs while continuing to rely heavily on Amazon. Google therefore has substantial internal demand and real external customers.

Its weakness is distribution. Nvidia hardware runs across every major cloud and reaches companies that do not want to depend too heavily on one provider. A TPU customer usually accepts a closer relationship with Google Cloud and its software environment.

Google can already operate much of its own AI business with limited Nvidia dependence. Becoming the neutral computing platform used across the wider industry would require a broader sales and software strategy than Google has shown so far.

Q6Is Amazon already replacing Nvidia inside AWS?

Amazon is already moving a meaningful amount of AI work away from Nvidia inside AWS.

Trainium has crossed a line that most custom-chip projects never reach: enormous production use. Amazon says 1.4 million Trainium2 chips have landed and the available supply is fully subscribed. Project Rainier contains roughly half a million processors, while Anthropic reports using more than one million Trainium2 chips across its broader AWS infrastructure to train and serve Claude.

The relationship extends well beyond a normal cloud contract. Anthropic has committed more than $100 billion over ten years to AWS technologies and secured several gigawatts of additional capacity. Its engineers work directly with Amazon's Annapurna Labs team on the hardware and software. Amazon can improve Trainium around a small number of extremely valuable workloads instead of supporting every possible AI application immediately.

Amazon also says its custom chips now handle most inference on Bedrock, which is used by more than 100,000 companies. Trainium3 is already running production workloads, with nearly all planned capacity expected to be committed. Those numbers describe a recurring platform, not a promotional experiment.

AWS will continue offering Nvidia because customers still ask for CUDA and want to move models between clouds. Amazon's practical strategy is to place predictable, high-volume workloads on Trainium while selling Nvidia capacity whenever flexibility matters more.

Over time, Amazon can keep the repeatable workloads that are easiest to optimize and most attractive economically, leaving Nvidia with more of the unfamiliar and less standardized jobs.

Q7Can Microsoft and Meta replace Nvidia from inside?

Microsoft and Meta can move a substantial amount of their own AI work away from Nvidia, although neither is close to abandoning it.

Microsoft's Maia 200 was designed for inference and already operates inside its data centers. Microsoft says it delivers more than 30% better tokens per dollar than the latest hardware previously used in its fleet. The comparison comes from Microsoft rather than an independent benchmark, but the economic objective is clear: lower the recurring cost of serving stable models at Azure scale.

Microsoft is widening its supplier list rather than choosing one replacement. Azure supports Nvidia, Microsoft's own Maia chips and a growing AMD fleet. The latest AMD agreement expands that fleet again, while Microsoft continues preparing Azure for Nvidia's Vera Rubin systems. The company wants to match different hardware to different workloads.

Meta is following a similar strategy with more internal silicon. It plans four new generations of MTIA chips within two years for recommendations, ranking and generative AI. Broadcom is helping develop several of those generations, and Meta is also working with Arm on custom data-center processors. Alongside that roadmap, Meta has signed large agreements with Nvidia and AMD.

Neither company expects one architecture to handle all its AI work efficiently. Internal chips will absorb stable workloads they understand well. General-purpose accelerators will remain useful for rapidly changing models and for outside customers.

Both companies can reduce Nvidia's share of their infrastructure spending. Neither currently offers an internal platform that independent developers would adopt as a broad alternative to CUDA.

Q8Is Broadcom quietly building the post-Nvidia market?

Broadcom may be doing more damage to Nvidia than any company without a famous rival GPU.

Its strength comes from helping large customers design chips around their own workloads. Broadcom provides semiconductor engineering, packaging, connectivity and networking expertise. The customer contributes its models, traffic patterns and economic incentive to avoid Nvidia's margins.

This business has become far too large to dismiss as a side project. Broadcom's latest quarterly AI semiconductor revenue reached $10.8 billion, up 143% from the previous year. Management expects about $16 billion in the following quarter, which would represent annual growth above 200%.

OpenAI offers the clearest example. It plans to deploy 10 gigawatts of OpenAI-designed accelerators with Broadcom by 2029. The companies have also introduced Jalapeño, an inference chip intended to begin deployment before the end of 2026. Meta has separately selected Broadcom to help develop several generations of MTIA.

Broadcom does not need to create one universal accelerator. It can help OpenAI, Meta and other major buyers build different chips, then provide networking and connectivity across several of those systems. The market could fragment across multiple architectures, reducing Nvidia's volume without producing one obvious successor to CUDA.

Broadcom is the leading candidate to equip a customer-led attack. The hyperscalers would own the workloads and accelerators; Broadcom would supply much of the engineering machinery behind them.

Q9Can Huawei push Nvidia out of China?

Huawei can push Nvidia out of much of China even if Ascend never wins the global market.

China is a special battlefield because availability and political support can outweigh a clean benchmark comparison. Nvidia's latest outlook assumes no data-center compute revenue from China, while its regulatory filings describe official pressure on Chinese buyers to choose domestic suppliers. Export controls have already forced Nvidia to redesign products and absorb billions of dollars in inventory-related charges.

Huawei now sells more than an isolated accelerator. The Atlas 900 A3 SuperPoD connects as many as 384 Ascend chips into one system, supported by Huawei's CANN software, development tools and local engineering teams. Huawei says the platform is being used by internet, finance, telecommunications and power-sector customers.

Chinese companies also face a planning problem that favors Huawei. An Nvidia product may be technically attractive and still become unavailable after another policy change. A domestic roadmap gives buyers more certainty when they are building data centers expected to operate for years.

Huawei's global reach remains restricted by weaker software, sanctions and limited access to advanced manufacturing. Those disadvantages carry less weight inside China because Nvidia's own access is constrained. The likely result is a large parallel Chinese ecosystem rather than a single worldwide winner.

Q10Can Cerebras, Groq or SambaNova take Nvidia's best workloads?

Specialist chipmakers can capture profitable parts of Nvidia's business, but none currently supports enough workloads to replace its platform.

Cerebras has the strongest evidence of large-scale adoption. Its wafer-scale architecture places enormous computing capacity, memory and bandwidth on one giant chip, reducing some of the communication bottlenecks found in conventional clusters. OpenAI has agreed to add 750 megawatts of Cerebras capacity for low-latency inference, and the partnership is already being used for a real-time coding model producing more than 1,000 tokens per second.

Groq concentrates on predictable, extremely fast inference. SambaNova aims its dataflow architecture at agentic workloads and large-model serving. Both can be attractive when latency and throughput matter more than compatibility with every framework or model architecture.

Their opportunity remains narrow. A specialist can deliver an exceptional result on one workload while being awkward for research, training or models outside its target area. Nvidia spreads its software and engineering investment across a much broader market.

Where specialist AI chips can attack Nvidia

Q11Is inference where Nvidia is easiest to beat?

Inference is where Nvidia currently looks easiest to beat.

Frontier-model training changes quickly. Researchers experiment with new architectures, kernels, numerical formats and communication patterns, which rewards the flexibility and mature tooling of a general-purpose GPU. Serving the same model millions or billions of times creates a steadier workload. Once the traffic becomes predictable, a custom accelerator can be designed around it.

We see that preference across companies with very different strategies. Microsoft's Maia 200 focuses on inference economics. Amazon says its custom chips handle most Bedrock inference. OpenAI designed its new Broadcom chip for serving models and selected Cerebras for low-latency output. Google's latest TPU covers training and inference, with some of its largest claimed economic improvements coming from inference workloads.

Inference also expands with usage. A model may be trained a limited number of times, but every query, generated image, recommendation and agent action consumes serving capacity. As AI products gain users, that recurring bill can eventually exceed the original training cost.

Nvidia is responding aggressively through lower-precision formats, additional memory, faster networking and rack systems designed around token throughput. It still has a strong position. The problem is that inference has become large and repetitive enough for customers to justify their own silicon, making Nvidia's broad flexibility less valuable for certain jobs.

Q12Can open software finally loosen CUDA's grip?

Open software is loosening CUDA's grip, although Nvidia still provides the safest path for unfamiliar workloads.

CUDA's advantage extends far beyond the programming language. It includes optimized libraries, debugging tools, compilers, communication software, documentation and a huge pool of engineers who already know how to make Nvidia clusters perform. That accumulated experience saves time during extremely expensive model runs.

The alternatives have improved. ROCm now covers compilers, runtimes and important libraries, with direct support in PyTorch. AMD's HIP tools can translate and maintain much CUDA-style code. Google has JAX and XLA, Amazon has the Neuron SDK, Huawei has CANN, and several frameworks increasingly hide hardware-specific work behind compilers and standard operators.

Portability remains weakest where frontier teams care most. New attention methods, quantization techniques and custom kernels often require manual tuning. Code may technically run on another accelerator while using it inefficiently, which can erase the saving on hardware.

CUDA will have truly lost its grip when teams can optimize the model first and select the hardware afterward. Today, many teams still choose the hardware early because the available software shapes what they can build efficiently.

Q13Can open networking break Nvidia's rack advantage?

Open networking can weaken Nvidia's rack advantage, but the alternative ecosystem is still being assembled.

A modern AI cluster depends heavily on the connections between chips. Thousands of accelerators must exchange data, reach shared memory and recover from failures without spending most of their time waiting. Nvidia addresses this through NVLink, InfiniBand, Spectrum-X, ConnectX and BlueField, all designed to work closely with its GPUs and rack systems.

The rival camp is becoming more organized. The UALink consortium includes major cloud providers, chipmakers and system suppliers. It has published technical specifications, while participating companies are developing compatible switches, retimers, connectors and transport systems.

That gives AMD and custom-chip developers a chance to build large clusters without relying entirely on Nvidia's proprietary fabric. Customers would also gain more freedom to combine processors, switches and memory from different suppliers.

Open standards still have to become reliable products. Nvidia ships working systems today; the alternatives need hardware, software, validation and years of field experience. Nvidia has already responded with NVLink Fusion, which allows outside CPUs and custom accelerators to connect to parts of its infrastructure. Its goal is to remain inside the rack even when another company supplies the main accelerator.

Q14Why doesn't a cheaper AI chip automatically beat Nvidia?

A cheaper AI chip wins only when the complete system produces useful work for less money.

Buyers pay for electricity, cooling, memory, networking, engineers and data-center space as well as the accelerator. Delays cost money too. A chip with a lower purchase price can become expensive when software takes months to port, the cluster operates below capacity or failures interrupt a large training run.

The figures that count are cost per completed training run and cost per usable token. Microsoft says its internal accelerator cuts token costs by more than 30% compared with the previous hardware in its fleet. Amazon markets Trainium around similar economics. Such savings are plausible inside tightly controlled workloads because the company owns the chip, software, model and data center.

An independent startup faces a different calculation. It may still prefer Nvidia because its engineers already know CUDA, its model can run across several clouds and its future workloads are difficult to predict. Flexibility can be worth more than a lower chip price.

Nvidia's latest gross margin was about 75%, giving customers a powerful incentive to search for alternatives. The first damage will probably appear in pricing and workload selection. Customers can keep Nvidia for jobs that benefit from its platform while moving stable volume elsewhere.

Q15Are Nvidia's biggest customers its biggest danger?

Nvidia's biggest customers are currently its greatest long-term danger.

Its latest filing shows that three direct customers represented 21%, 17% and 16% of quarterly revenue. Those categories can include cloud providers, server manufacturers, distributors and model developers, so the filing does not reveal the final buyers. It still shows remarkable concentration: three direct accounts produced 54% of company revenue.

The same group has unprecedented resources to build alternatives. Based on their latest guidance, Amazon, Alphabet, Microsoft and Meta plan roughly $695 billion to $725 billion of combined capital spending in 2026. Not all of that money will go toward AI chips. Amazon also funds logistics and satellites, while every company invests in buildings, networks and other equipment. Even after accounting for that, the resources available for AI infrastructure remain enormous.

Their incentive grows with every Nvidia purchase. Paying high margins makes sense while Nvidia saves years of engineering work. Once a workload becomes stable and large enough, spending several billion dollars on internal silicon can reduce a much larger recurring bill.

These customers do not need to sell their accelerators to the broader market. Google can justify a TPU through Gemini and Google Cloud. Amazon can justify Trainium through AWS and Anthropic. Microsoft and Meta can target their own largest inference and recommendation workloads. They need internal savings; an independent chip startup must create a sales channel and persuade outside developers to switch.

How Nvidia's largest buyers are building alternatives

Q16Is Nvidia adapting fast enough?

Nvidia is adapting quickly enough to remain the leader for now.

The company now sells much more than accelerators. It increasingly sells the design of the entire AI data center. Vera Rubin combines GPUs, a purpose-built CPU, networking and storage acceleration. Nvidia's custom-chip interconnect program gives outside processors a route into its network. DGX Cloud software aims to keep developers in a familiar environment even when the physical machines sit inside different clouds.

Its new reporting structure provides another useful clue. Nvidia now separates data-center revenue into Hyperscale and a second group covering AI clouds, industry and enterprise. The two were almost equal in the latest quarter, at roughly $37.9 billion and $37.4 billion. Its growth therefore extends beyond a small group of hyperscalers into sovereign projects, industrial systems and specialized AI clouds.

That broader reach is difficult for internal chips to reproduce. Amazon can sell Trainium through AWS, but a government laboratory, carmaker or robotics company may prefer technology available from several suppliers and supported by a much larger developer community.

Nvidia is also turning rival strategies into products it can sell. Custom accelerators can connect through NVLink Fusion. Customers moving toward Ethernet can buy Spectrum-X. The growing importance of inference creates demand for new Nvidia systems tuned around token output and lower numerical precision.

This makes a sudden collapse unlikely. Nvidia can lose some accelerator volume while protecting revenue through networking, CPUs, software and complete systems.

Q17What would Nvidia have to get wrong?

Nvidia would need to make several expensive mistakes at once before anyone could truly dislodge it.

The most dangerous error would be pushing prices beyond the value of convenience. High margins finance Nvidia's lead, but they also strengthen every customer's custom-chip business case. When a stable workload can save billions of dollars over several years, the cost and difficulty of leaving CUDA become easier to accept.

A second error would be mishandling the shift toward inference. Nvidia built its current position during an era when flexible GPUs were ideal for rapidly changing training workloads. A market increasingly dominated by repeatable model serving gives Google, Amazon, Microsoft, OpenAI and specialist companies more room to optimize around narrower jobs.

Software could become another weakness if developers started viewing CUDA mainly as a lock-in mechanism. Nvidia has to keep earning loyalty through better performance and productivity. Aggressive restrictions or poor portability would push more money and talent toward open compilers and hardware-neutral frameworks.

Execution matters as well. One generation affected by poor yields, cooling problems, networking failures or long deployment delays could redirect billions of dollars, especially now that credible alternatives exist. Geopolitics creates a separate risk because China can develop a large domestic ecosystem that Nvidia may never fully recover.

One setback would hurt. The real danger is several landing together: excessive pricing, a missed workload shift, declining developer trust and a badly executed product cycle.

Q18Who can eventually kill Nvidia?

No single company is likely to kill Nvidia; a customer-led coalition has the only credible path.

Google is the strongest standalone challenger because it already owns a complete model-to-chip stack. AMD is the best direct substitute for buyers that still want a widely available general-purpose accelerator. Broadcom is the key enabler, helping large customers turn their workloads into custom silicon. Amazon has shown the greatest disclosed scale for an internally controlled cloud accelerator. Huawei can remove Nvidia from much of China. Cerebras and other specialists can capture narrow but valuable parts of inference.

Every contender also has a serious weakness. Google's TPU remains closely tied to Google Cloud. Amazon's Trainium is an AWS product. AMD is much smaller and still catching up in software and complete systems. Broadcom enables other companies rather than providing a common developer platform. Huawei faces manufacturing and geopolitical limits outside China. Specialist chips cover only part of the market.

Their strengths fit together. Custom silicon attacks Nvidia's margins. AMD provides a second general-purpose option. Open software lowers the cost of moving code. UALink and other standards give rival accelerators a route into large systems. Specialist processors take repeatable inference workloads. Huawei builds a separate Chinese ecosystem.

Nvidia should remain the largest AI infrastructure supplier for years. Its revenue is still surging, its product range is widening and its customers continue signing large Nvidia agreements even while developing alternatives. A near-term defeat would ignore the extraordinary scale of its current lead.

The long-term threat is real. Nvidia loses its exceptional position once its largest customers no longer need one company to provide the default accelerator, software and network. Broadcom may build much of the machinery, AMD may win the most visible market share, and Google may come closest to matching the complete stack. The group capable of pulling it off is the same group currently sending Nvidia most of its money.