AI, particularly the training of generative AI large language models (LLMs), continues to drive the majority of semiconductor market activity. NVIDIA Corp.'s launch of its Blackwell graphics processing unit (GPU) in March at the company's GPU Technology Conference (GTC) was the centerpiece and attracted record crowds of developers, data scientists, GPU cloud clustering vendors and systems makers to San Jose, Calif.

Also in March, Intel Corp. landed the largest ($8.5 billion) funding deal yet under the US CHIPS Act, and revealed the latest plans in its foundry services roadmap, while China announced that it would be raising $27 billion to support its domestic chip industry. China also disclosed that it would no longer approve Advanced Micro Devices Inc. and Intel chips for deployment in government PCs and servers.

The new CHIPS Act funding for Intel — the first major award since the act first came into being in 2022 — was broadly welcomed by the industry. However, it is not yet clear if the entire $35 billion CHIPS Act funds originally allocated for spending by 2025 will be awarded in time, or whether further investments will be made beyond that date. Many of the semiconductor production plants planned for the US have been focused on chip production rather than packaging, leaving a risk that US semiconductor independence could be constrained by a lack of downstream capabilities.

Some caution is needed, though, as plants planned by Taiwan Semiconductor Manufacturing Co. Ltd. and Intel have been delayed due to late payment of subsidies and labor availability. A long-term sustained investment program will be necessary if the goals of the act are to be achieved, among them supply chain resiliency, national security and the availability of skills to repatriate manufacturing to the US.

Product releases, updates

Intel revealed more of its foundry plans at the end of February, calling Intel Foundry "the world's first system foundry for the AI era." Its roadmap includes a new Intel 14A process technology (joining Intel 7, 4, 3, 20A, 18A and 16), specialist node evolutions, and advanced system assembly and test capabilities. 14A will be the company's smallest node yet deploying High NA EUV lithography. Microsoft Corp. is a new design win, with a chip design that will be built with Intel's 18A process technology. Synopsys Inc., Cadence Design Systems Inc., Siemens Ltd. and ANSYS Inc. have announced support for Intel 18A and advanced packaging in their tools. Intel's overall plans include a new node every two years, with node evolutions in between.

Arm Holdings PLC introduced its automotive-enhanced family of processor designs for carmakers, along with new Compute Subsystem components and a software stack aimed at reducing development time and cost. It enables virtual prototyping for the first time.

Early in March, Hewlett Packard Enterprise Co. said during its first-quarter results call that the GPU shortage reduced its sales of AI servers. The systems-maker is seeing GPU lead times of about 20-plus weeks. Also, customers are taking longer to set up the surrounding Infrastructure, including power and cooling, for AI servers. HPE has accumulated a $3 billion order backlog for AI servers. Dell Technologies Inc. saw orders for AI servers increase 40% sequentially from third quarter 2023 to first quarter 2024, with its backlog doubling. Lenovo Group Ltd. also saw an uptick and expects AI server sales to grow nearly twice as fast as the total server market. Super Micro Computer Inc. claims that in the second quarter of 2024, half of its total revenue of $3.66 billion came from AI.

Cerebras has unveiled what it claims to be the world's fastest AI chip. The Wafer Scale Engine 3, using up to 4 trillion transistors on a 5-nanometer process technology, is capable of training LLMs up to 10 times larger than the current GPT-4 or Gemini LLMs. The chip powers the 125-petaflop Cerebras CS3 supercomputer, with 900,000 AI cores on the same processor. Cerebras has a co-design partnership with QUALCOMM Inc. involving the latter's AI 100 Ultra chips for inference.

At the GTC conference in San Jose, which featured a stadium-scale keynote by CEO Jensen Huang, Nvidia introduced its new Blackwell family of datacenter GPUs, successors to the previous flagship H100 GPU family. There are several flavors: the B100, the B200 and the combined CPU-GPU GB200 Grace-Blackwell superchip. A shared memory rack with 72 Blackwell nodes and 36 Grace CPUs was also displayed — and eight of these can be linked to make the latest generation Superpod. A software highlight was the launch of Nvidia Inference Microservices for running pretrained AI models with all of their dependencies containerized, making them easily portable across clouds, datacenters, workstations and PCs.

Qualcomm launched its Snapdragon 8 Gen 3 system-on-chip (SoC) at Mobile World Congress in February, enabling on-device AI for devices, including the iPhone and the Samsung Galaxy S24 smartphone. Later this year, Qualcomm plans to ship its Snapdragon Elite X, combining CPU, GPU and neural processing unit in an SoC aimed at future AI PCs.

M&A activity

Consolidation in the chip sector continues to be curtailed by underlying geopolitics: Chinese attempts to acquire UK- and Germany-based semiconductor providers have been blocked, while China scuttled Intel's proposed $5.4 billion acquisition of Tower Semiconductor Ltd. by withholding regulatory approval. Although deals in the semiconductor sector continue to be few and far between, the start of 2024 saw Synopsys Inc. print its largest-ever acquisition with the purchase of Ansys Inc. for $35 billion. That transaction moves Synopsys beyond its traditional electronic design automation (EDA) business into the broader simulation space, just as the worlds of simulation and AI modeling move closer together. Meanwhile, Renesas Electronics Corp. is in the process of buying two companies: power management chip specialist Transphorm Inc. for $339 million and EDA tools maker Altium Ltd. for $5.7 billion.

Funding and further consolidation

As for funding, new chip startups (mostly AI-focused) continue to emerge, but those in the US and Europe are starting life with the prospect of having no access to the largest market in the world.

In March, rumors emerged that Nvidia was in advanced talks to acquire Run.ai, an Israel-based startup that offers a containerized platform for virtualizing and pooling GPU resources, from fractions of a single GPU to large clusters. Run.ai has raised over $105 million in funding since launching in 2019.

Meanwhile, AI chipmaker Groq paid an undisclosed amount for Definitive to bolster its cloud unit, GroqCloud. The target has developed natural language tools for querying datasets. Groq has been moving into cloud services as a means of exposing its Language Processing Unit (LPU) inference engine to customers, optimizing its infrastructure for specific open-source models such as Mistral and Meta Platforms Inc.'s Llama 2.

The UK has rejoined Horizon Europe as a participating state, offering Britain-based startups access to the €1.3 billion research fund for the next four years and giving the country a role in setting future priorities for research areas. Current focus areas include automotive processors and the development of open-source RISC-V designs.

Astera Labs Inc. filed for its IPO on the Nasdaq in late March, with its share price increasing almost 70% by the end of the day, and raising $712.8 million, some $200 million above its expectations. Astera sells chips, boards and modules for high-speed connectivity between chips, systems in a rack, or racks. The vendor was early to market with devices supporting the Compute Express Link protocol.

Chiplet interconnect startup Eliyan has raised $60 million in a series B funding round led by Samsung Catalyst Fund and Tiger Global Management, with participation from Intel Capital and SK hynix Inc. The tranche follows a $40 million series A in 2022. Eliyan is working on NuLink, a die-to-die interconnect for chiplets, as well as a bidirectional Universal Memory Interface for aggregating memory bandwidth.

South Korea's Rebellions has raised $124 million in a series B funding round led by KT Corp., Pavilion Capital and Shinhan Ventures for the development of its "next generation" AI chip for running large language models.

AI chip specialist Recogni has raised $102 million in a series C funding round led by Celesta Capital and GreatPoint Ventures. The company is working on AI inference silicon for generative AI and intelligent autonomy in the automotive and aerospace sectors.

SEMRON, a chipmaker headquartered in Germany, has raised €7.3 million in seed funding to develop AI chips for mobile devices. The chips are focused on cost, small form factor and energy consumption.

Toronto-based Taalas is developing a platform that can quickly turn any AI model into custom silicon to create "hardcore models" that it claims will be 1,000 times more efficient than their software counterparts. The vendor has come out of stealth and says it has raised $100 million in funding from Pierre Lamond and Quiet Capita. Its founder is Ljubisa Bajic, who previously founded AI accelerator company Tenstorrent. Taalas notes that its first processor will be announced later this year, with availability in 2025.

Geopolitics, regulations

Intel has been awarded $8.5 billion in grants to expand its manufacturing capabilities in the US, the largest award under the CHIPS Act so far. The deal also includes up to $11 billion in loans and tax credits. The vendor is building new plants in Arizona, New Mexico, Ohio and Oregon. The US hopes to see about 20% of the world's leading-edge chips produced in the country by the end of the decade. Other recent awards under the CHIPS Act have gone to BAE Systems PLC ($35 million), GLOBALFOUNDRIES Inc. ($1.5 billion) and Microchip Technology Inc. ($162 million) to bolster the production of legacy chips.

The UXL Foundation (the Unified Acceleration Foundation), a consortium of chipmakers looking to establish an alternative to Nvidia's CUDA software stack that will run atop multiple processors, is reported to be readying its first detailed specifications in the first half of this year. They will be based on SYCL and Intel's OneAPI framework, including technology that the latter acquired from startup Codeplay Software. Foundation members include Arm, Broadcom Inc./VMware, Fujitsu Ltd., Google Cloud, Imagination, Intel, Qualcomm and Samsung Electronics Co. Ltd.

China is aiming to raise over $27 billion as part of the National Integrated Circuit Industry Investment Fund, which is now entering its third phase. The so-called "Big Fund" is intended to counteract tightening US export controls. The third phase is particularly focused on semiconductor manufacturing equipment for producing legacy chips, where China hopes that it can maintain supply dominance and pricing advantages.

Meanwhile, Huawei and Semiconductor Manufacturing International Corp. (SMIC) are planning to collaborate on the production of 5-nanometer processors in Shanghai. Last August, Huawei's HiSilicon division supplied a more advanced than expected Kirin 9000 chip using 7-nanometer process technology. The chip is intended for deployment in the Huawei Mate 60 Pro smartphone.

Additionally, China's Information Security Evaluation Center has removed CPUs by Intel and AMD from its approved list for use within government PCs and servers, recommending in their place alternatives from Loongson Technology Corporation Ltd., Phytium Technology and Shanghai Zhaoxin Semiconductor — the latter being the only x86 option. The country is also restricting the deployment of US operating systems (including Windows) and databases.

^{This article was published by S&P Global Market Intelligence and not by S&P Global Ratings, which is a separately managed division of S&P Global.
451 Research is a technology research group within S&P Global Market Intelligence. For more about the group, please refer to the 451 Research overview and contact page.}