Between August 2025 and July 2026 the chip business stopped revolving around the fastest processor alone. Nvidia stretched into inference and software, Intel looked for capital and partners to sustain its manufacturing bet, AMD and Qualcomm pressed from the flanks, and export controls pushed China to build its own technology stack.
Signed by Falcon, Signals analystUpdated on What happened and when
Washington weighs a stake in Intel
The US government was reported to be considering a stake in Intel after Trump criticised its chief executive; the company dismissed it as rumour. SoftBank announced a $2 billion investment at $23 a share.
Jetson Thor goes on sale
Nvidia launched its Jetson AGX Thor robotics module from $3,499, which it says is 7.5 times faster than the previous generation, and named robotics its biggest opportunity outside AI.
China accuses Nvidia of antitrust breach
China's regulator preliminarily found that Nvidia failed to meet the conditions attached to approval of the Mellanox purchase, amid a trade standoff between Washington and Beijing.
Nvidia takes a stake in Intel
The deal includes a $5 billion investment and RTX graphics chiplets for Intel processors. Intel said it complements its roadmap and that it will keep making GPUs.
Qualcomm buys Arduino and targets the data centre
Qualcomm announced the Arduino purchase to reach people who prototype robots and unveiled its AI200 and AI250 accelerators, competing with Nvidia and AMD in rack-scale systems.
Lawsuits over chips in Russian weapons
Ukrainian civilians sued Texas Instruments, AMD and Intel in Texas for failing to track where their chips ended up, which the plaintiffs say was Russian and Iranian weaponry.
Nvidia licenses Groq as Intel backs 14A
The $20 billion Groq licence was read as the end of the general-purpose GPU as the single answer for inference. Intel reaffirmed its commitment to the 14A process.
GTC brings Vera Rubin and NemoClaw
Nvidia presented Vera Rubin, a seven-chip platform, and NemoClaw for enterprise agents. It arrived after quarterly revenue of $68.1 billion, up 73% on a year earlier.
Nvidia invests $2 billion in Nebius
The investment comes with a deal for Nebius to deploy more than 5 gigawatts of Nvidia-based capacity by the end of 2030, with early access to Rubin.
Super Micro co-founder charged
US prosecutors charge him and two others with diverting to China servers carrying Nvidia chips worth about $2.5 billion through a shell company.
Agent Toolkit signs up 17 partners
Nvidia launched an open-source agent toolkit that 17 software companies, including Adobe, Salesforce and SAP, committed to adopt.
B300 servers reach $1 million in China
Scarcity caused by export controls pushed B300 server prices to around a million dollars. Huang warned of the risk of DeepSeek moving to Huawei chips.
Cerebras debuts as Alphabet closes in on Nvidia
Cerebras nearly doubled on its first day of trading, reaching a $100 billion valuation. Alphabet passed $4.6 trillion in market value and narrowed the gap with Nvidia.
A Chinese model trained on domestic chips
Meituan says LongCat-2.0 was trained and runs end to end on a cluster of 50,000 domestic chips.
Nvidia forecasts trillions in AI spending
The company expects AI infrastructure investment to reach between $3 trillion and $4 trillion by 2030.
The threads that matter
Nvidia: from chip supplier to full platform
Nvidia spent the period widening its perimeter. First came the Jetson AGX Thor robotics module, from $3,499 and, by the company's own figure, 7.5 times faster than the previous generation; then the $20 billion licence of Groq's technology, which one analysis read as the beginning of the end for the general-purpose GPU as the single answer for inference. In March 2026 Vera Rubin gave that idea shape: seven chips, including a Groq-based inference accelerator, grouped into rack-scale systems.
The second front was software. NemoClaw was presented as an enterprise layer that adds security and scale to autonomous agents, and the Agent Toolkit gathered 17 software companies, among them Adobe, Salesforce and SAP. Everything is open source yet optimised for Nvidia hardware, and one analysis reads this as a way for every new agent to generate GPU demand; it is also a turn away from the company's historic dependence on CUDA.
The open Nemotron models complete the picture. Nemotron 3 Super combines three architectures in 120 billion parameters; Cascade 2 wins gold medals with only 3 billion active parameters and beats a model with four times as many; and Nano Omni unifies vision, audio and language. The common message is that efficiency matters more than raw size.
- Nvidia Lanza el Cerebro Robotico Jetson AGX Thor Potencia IA para Robots y Mas
- GPUs de Proposito General El Fin de una Era
- Nvidias Vera Rubin Revolutionary Seven-Chip AI Platform
- Nvidia Desata Poderosas Claws IA Empresariales Seguras
- Nvidias Agent Toolkit Wins Over 17 Enterprise Giants
- Nvidia Lanza NemoClaw Plataforma de Agentes IA
Intel: capital, partners and a bet on manufacturing
Intel spent the year looking for backing. In August 2025 the US government was reported to be weighing a stake after Trump criticised chief executive Lip-Bu Tan, something the company dismissed as rumour; the background included $8 billion in CHIPS Act funding and a $28 billion complex in Ohio. Shortly afterwards SoftBank announced a $2 billion investment at $23 a share, after a year in which the shares lost 60% of their value.
The most unexpected ally was Nvidia, which invested $5 billion and supplied RTX graphics chiplets for Intel processors. Intel insisted the deal complements its roadmap. In January 2026 its chief executive reaffirmed the bet on the 14A process, which unlike 18A could attract a significant external customer. Separately, Intel quietly stopped maintaining the open-source driver for its Gaudi accelerators, hosted on a code-hosting platform, and encouraged others to fork it.
- US Government Stake in Intel Trump Considers Deal After CEO Attack - What It Means for Chip Ma...
- SoftBank invierte 2 mil millones en Intel Impulso para el Gigante de los Chips en la Era de la IA
- Intel ha respondido a las especulaciones sobre el impacto del reciente acuerdo con Nvidia en sus…
- Intel Apuesta Fuerte por el Proceso 14A CEO Confirma planes
- Intel Deja el Codigo Abierto de Gaudi Que Significa
The challengers: Qualcomm, Cerebras, TensorWave and AMD
With Nvidia dominant, several players looked for an opening. Qualcomm unveiled the AI200 and AI250 data-centre accelerators, designed for liquid-cooled rack systems like those Nvidia and AMD already offer, which let up to 72 chips act as one computer, and bought Arduino to get closer to people who prototype robots. TensorWave, a cloud specialised in AMD hardware, raised $350 million in a round led by AMD and runs about 8,000 Instinct MI325X accelerators, while Cerebras nearly doubled on its first day of trading, at a $100 billion market value.
AMD mixed technical progress with trust problems. It worked with Sony on the Project Amethyst architecture, and its hardware let IBM run quantum error correction in real time. But a flaw in the SEV-SNP confidential-computing feature allowed a virtual machine to be compromised with a single 8-byte write, and the company had to backtrack twice after user pushback: first on game support for older Radeon cards, then by restoring the memory encryption it had removed through firmware.
China, export controls and self-reliance
Trade tension shaped the Chinese market. In September 2025 the regulator preliminarily found that Nvidia breached the conditions of its Mellanox purchase, made in 2020 for $6.9 billion, with possible fines of between 1% and 10% of the previous year's sales. By April 2026 B300 servers were selling in China at around a million dollars, nearly double the US price, because of the scarcity the controls cause. And in May it was reported that Huang said Nvidia had lost all its share there and that export policy had failed.
Enforcement also reached the courts. A Super Micro co-founder was charged with diverting to China servers carrying about $2.5 billion of Nvidia chips, and Ukrainian civilians sued Texas Instruments, AMD and Intel in Texas for not policing where their chips ended up. Meanwhile China is leaning toward application-specific chips (ASICs): Huawei is projected to reach 62% of domestic accelerators in 2026, and Meituan says it trained a 1.6-trillion-parameter model on 50,000 domestic chips. Huang warned it would be a horrible outcome if DeepSeek optimised for Huawei.
- China Acusa a Nvidia Violacion Antimonopolio
- Nvidia B300 Servers Hit 1M in China AI Chip Scarcity Soars
- Nvidia sin cuota en China Exportaciones de EE UU fracasan
- Super Micro Founder Charged 25B GPU Smuggling to China
- US Chip Firms Sued Ukraine War Weaponry Allegations
- China IA Sin GPUs NVIDIA Nueva Era Tecnologica
Why it matters and what to watch
From selling chips to selling platforms
If Nvidia's open-source agents become widespread, competitive advantage shifts from the silicon to the ecosystem around it; watch how many partners move from adopting the toolkit to depending on it.
Inference as the new battleground
Groq, Qualcomm and Cerebras all target the same bottleneck, latency in responding; the figure to follow is whether specialised accelerators take share from GPUs in real workloads.
Supply chains and industrial policy
Export controls create scarcity, grey markets and lawsuits, and push China to replace both hardware and software; each new restriction accelerates that split.
Capital and concentration
Nvidia projects AI infrastructure spending of between $3 trillion and $4 trillion by 2030 while Alphabet closes in on its market value; the risk is that so much capital depends on a handful of suppliers.
What we think
From the signals desk, the pattern is plain: the chip is now the entry point, not the product. Whoever controls the software, the models and the complete systems decides how costly it is to switch supplier, which explains Nvidia's insistence on opening code that is nonetheless optimised for its own hardware. We read the Groq licence as an admission that one architecture will not cover every workload, and the Chinese push for domestic silicon as a structural shift rather than a passing reaction. We are sceptical of performance claims that are not verified on real workloads; laboratory figures are an invitation to test, not a conclusion. The sensible watch-list is short: inference latency, software lock-in and the export rules.
Falcon, Signals analystEverything we published on this topic
43 stories from Aug 2025 to Jul 2026, by month. These are the original articles this guide rests on; each one links to its source.
