Tracking the market for semiconductors, from the fab floor to the customer’s door.

Date: June 2024
Editor and Author: David MacQueen
Lead Author: James Sanders

The June 2024 edition of the Chip Observer covers the latest tech trends, including Qualcomm’s Snapdragon X SoCs with a 45 TOPS NPU, AMD’s Ryzen AI 300 SoC with 50 TOPS, and Intel’s upcoming Lunar Lake SoCs with 48 TOPS NPUs. It highlights new AI accelerators from AMD and Intel, growing competition from custom ASICs, and a recovering advanced memory market. The issue also addresses increased tariffs on Chinese EVs and the delay of IPOs by twenty-three Chinese semiconductor firms due to trade restrictions.

Industry Observatory

  • Qualcomm’s Snapdragon X Arm-based Windows SoCs include an NPU with 45 TOPS.
  • Snapdragon X PCs, including Microsoft’s own Surface devices, will feature Microsoft’s Copilot+ PC branding, ahead of x86 processors.
  • At Computex, AMD announced its new x86 Ryzen AI 300 SoC with an NPU capable of delivering 50 TOPS, to ship in Q2.
  • Intel will follow in Q3 with its Lunar Lake architecture SoCs, which feature NPUs capable of 48 TOPS.
AI in Datacenters
  • AMD and Intel launched their next generation of AI accelerators, the Instinct MI325X and Gaudi 3, respectively.
  • Despite these new launches, NVIDIA’s biggest competition are custom ASICs; Google was the third-largest designer of datacenter processors in 2023, without selling a single chip.
  • Advanced memory markets continue to rebound driven by further AI cloud investments.
China economic decoupling
  • The European Commission increased tariffs of Chinese electric vehicles (EVs) from 10% to a variable rate of 17.4%–38.1%. This follows US measures, where tariffs will increase from 25% to 100% later this year.
  • Twenty-three Chinese semiconductor firms pulled IPOs, signaling turmoil partly driven by increasing trade restrictions.

Under the Microscope

The Era of AI PCs

The first batch of Microsoft’s Copilot+ PCs reached store shelves in June, starting with the Microsoft Surface Pro 11 and Surface Laptop 7, as well as systems from Windows OEM partners Acer, ASUS, Dell, HP, Lenovo, and Samsung. These also represent a redoubled effort on Microsoft’s part to advance Windows on Arm CPUs; Qualcomm’s new Snapdragon X Elite and X Plus systems-on-chip (SoCs) include a neural processing unit (NPU) with 45 trillions of operations per second (TOPS) for artificial intelligence (AI) acceleration. Microsoft’s Copilot+ PC labeling requires at least 40 TOPS and a keyboard with a dedicated Copilot key in place of the menu key on traditional PC keyboards.

Qualcomm’s Snapdragon X launch represents the culmination of a strategy that started with the acquisition of Nuvia in 2021, which designed the low-power, high-performance Arm-compatible Oryon CPU core, replacing the reference core that Qualcomm licensed from Arm in its previous Snapdragon SoCs for Windows PCs. Qualcomm claims a 180% improvement in single-threaded performance on the Snapdragon X Elite compared to the previous generation Snapdragon 8cx Gen 3. Microsoft’s new Prism translation layer in Windows 11 24H2 bridges software compatibility for existing x86 and x86-64 apps to Arm, improving performance by up to 20% compared to previous Windows 11 releases. Vitally, the Snapdragon X series improves performance for traditional applications—not just AI—closing a gap that has hampered efforts to promote Windows on Arm for over a decade.

AMD surprised audiences at Computex with the announcement of its Ryzen AI 300 SoC. The processors, codenamed “Strix Point,” include the new Zen 5 core as well as a new NPU capable of delivering 50 TOPS of INT8 performance. Ryzen AI 300 CPUs are in production—and, in the hands of OEM partners—with retail availability expected in mid-July. Intel’s Lunar Lake generation of processors, announced the week prior to Computex, integrate NPUs with 48 TOPS and are expected in market in Q3 2024.

AI in Datacenters

Demand for AI training and inference capabilities in datacenters continues to move markets, particularly as NVIDIA faces competition from AMD, which is releasing its Instinct MI325X GPU, and Intel, which is releasing its Gaudi 3 accelerator this year, to compete against NVIDIA’s Blackwell GPUs. Likewise, NVIDIA’s cloud platform partners—Amazon, Meta, Microsoft, and Google—are developing custom AI accelerators to reduce (not eliminate) their reliance on NVIDIA GPUs. Adoption of Google’s TPUs in its cloud service propelled the company to become the third-largest designer of datacenter processors in 2023, without selling a single chip.

In turn, this demand for accelerated datacenters (and the marketing—if not yet demand—of AI PCs) is contributing to the recovery of the DRAM market. The launch of faster, denser HBM3e DRAM this year—used in NVIDIA’s Blackwell and AMD’s MI325X GPUs—will benefit AI performance and contribute to infrastructure upgrades and net-new purchases. Demand for high-bandwidth memory (HBM) is making supply scarce, with Micron and SK hynix’s production capacity sold out through 2024. Micron’s CEO indicated the “overwhelming majority” of supply for 2025 is already allocated.

Apple’s custom silicon efforts are extending into the datacenter, in support of new AI coming to macOS and iOS this fall. The initiative, called “Apple Chips in Datacenters” (ACDC) uses Apple Silicon—initially the M2 Ultra, with plans to upgrade to the M4 Ultra next year—for remote AI inference. Apple detailed plans to use secure enclaves in this context to ensure security of customer data, which could prompt wider adoption of secure enclaves (a security-in-silicon approach, found in recent datacenter-grade Intel and AMD CPUs, as well as the AWS Nitro System, a hardware hypervisor system) for AI contexts as well as general data processing.

China Decoupling from Global Tech Economy

In June, the European Commission confirmed an increase on tariffs of Chinese battery electric vehicles (BEVs) from a flat 10% to an additional 17.4%–38.1% depending on the cars’ attributes and the companies that produce them. These tariffs follow similar measures in the United States, with tariffs for Chinese EVs set to increase from 25% to 100% later this year. While this shields US and EU automakers from short-term competition in EVs, this is also likely to constrain supply chains and manufacturing strategies.

Shanghai Wusheng Semiconductor declared bankruptcy, and 23 other Chinese semiconductor firms have pulled their IPOs, signaling turmoil in the Chinese market driven in part by increasing trade restrictions. TechInsights’ G. Dan Hutcheson notes that “China’s typical economic policy has been not to pick winners, but to fertilize and water a market with far more money than it can absorb. This generates thousands of competitors in the initial phase, few of which ever generate the profitability to survive.” However, Hutcheson also notes that China risks the tactical mistake of developing its own set of standards—rather than following international standards—which will serve to isolate the country’s electronics industry, rather than protect them.

Data Observatory

TechInsights’ at-a-glance health check on the pulse of semiconductor manufacturing.

This chart shows trailing 12-month data for fabrication equipment sales versus integrated components revenues, overlaid against capacity utilization.

Data Observatory

  • Memory remains the hottest segment. HBM capacity is sold out to the end of 2025 at Micron and SK hynix.
  • Datacenter solid-state drives (SSDs) are selling fast, representing NAND’s first significant exposure to AI demand.
  • Semiconductor fab utilization has surpassed the 80% threshold, driven by leading-edge (5nm and better) processes. TechInsights anticipates it reaching an average of 90% in 2025.

TechInsights’ weekly TCI Graphics data stream offers market share and forecast data for the semiconductor industry. TCI Graphics brings you the power of nowcasting.

Editorial: Succeeding NVIDIA

NVIDIA briefly rose to the top of the stock market in June as the world’s most valuable company, worth $3.3 trillion at its peak. To the extent that stock prices can be interpreted as a validation of strategy, CEO Jensen Huang’s mission of transforming NVIDIA from a niche gaming silicon company to a datacenter technology powerhouse is complete. In 2023, NVIDIA’s datacenter GPU revenues totaled $36.3 billion—a 98% revenue share of that market, per TechInsights’ market sizing. In the battle of NVIDIA versus the rest of the semiconductor world, NVIDIA has won this round. There is no doubt that NVIDIA is succeeding—the question is, who might succeed NVIDIA?

Pretenders to the Throne

NVIDIA’s control of the market in 2023 reflected—in part—a lack of competitive products. When NVIDIA took a leap forward with Grace Hopper, competitors AMD and Intel were caught in the middle of two-year product refresh cycles. AMD’s competing GPU, the Instinct MI300, launched in December 2023. Intel’s Gaudi 3 AI accelerator started limited shipments in April 2024, with wider availably later in Q2.

AMD CEO Lisa Su is hoping to disrupt NVIDIA’s dominance in AI in the same way that AMD disrupted Intel’s dominance of datacenter CPUs. At Computex, Su announced a shift to a yearly cadence for datacenter GPUs, with a product roadmap for AMD Instinct that extends into 2026. AMD plans to refresh the Instinct MI300X using HBM3e RAM in the upcoming Instinct MI325X. This incremental mid-cycle refresh—a step rather than leap forward—is planned for Q4 2024.

Intel’s Gaudi 3 was the focus of its keynote at Computex. Intel took the unusual step of announcing the price of the product (rare in the world of datacenter silicon). We can assume that the $125,000 cost is below equivalent NVIDIA products. Intel is clearly positioning this generation not to compete on headline TOPS numbers but on cost effectiveness. Intel’s sales target—a rather modest $500 million, equating to a low single-digit percentage of the expected market in 2024—offered another acknowledgment that NVIDIA has won this round too. But Gaudi 3 is a precursor to Falcon Shores, Intel’s next-generation AI chip. Like AMD, Intel is also moving to an annual cadence, with Falcon Shores expected to debut in 2025.

While the new Gaudi 3 and MI325X are competitive against last year’s Hopper, NVIDIA has pushed ahead with its latest datacenter product, the Blackwell B200. It will be 2025 before AMD or Intel have a product competitive against NVIDIA’s latest and greatest, which is already in production. The shift to an annual refresh cadence by both AMD and Intel may ultimately put pressure on NVIDIA, but it’ll be another year at least until they have Jensen looking over the shoulder of his iconic leather jacket.

Startups are also pursuing the AI accelerator market. Cerebras, a company pursuing an architecturally differentiated “wafer-scale engine,” is reportedly planning an IPO this year. Tenstorrent, led by industry veteran Jim Keller, is another major startup in the market, as are SambaNova and D-Matrix. Designing and building processors is a capital-intensive task; so far no startup is providing significant competition to Intel or AMD, never mind putting pressure on NVIDIA.

ASICs: Putting the Custom in Customer

NVIDIA’s largest competitors are arguably its own customers. NVIDIA’s 98% market share for datacenter GPUs accounts only for the merchant silicon market. It does not include the custom chips, or application-specific integrated circuits (ASICs), major tech players are designing for use in their own datacenters. Companies such as Amazon, Google, Meta, and Microsoft may acquire some AI accelerators from NVIDIA but do not rely on one supplier. For the major tech players, typically the biggest other supplier is the company itself.

In 2023, Google was the second-largest producer of datacenter AI accelerators—and the third-largest designer of datacenter processors generally—without selling a single chip on the open market. Google’s first tensor processing unit (TPU), an AI accelerator ASIC, debuted in 2015. Its sixth-generation TPU was announced in May 2024, with availability planned later this year. Notably—and despite the two companies competing in other markets—Apple uses Google Cloud TPUs to train models for its Apple Intelligence functionality, which will be generally available later this year in select markets.

AWS introduced Inferentia, a custom AI inference ASIC, in 2018; Trainium, a custom AI training ASIC, launched in 2020. Both Amazon ASICs are now in their second generation. Meta develops custom ASICs for its internal AI workloads. The first iteration of its MTIA inference accelerator was launched in 2020, and next year the second generation should be available. Microsoft is a new entrant to AI ASICs, having announced in November 2023 plans to design custom silicon for AI for use in its Azure cloud platform. The initial version of that chip—the Maia 100—is used internally by Microsoft to run AI workloads on Azure, with general availability expected this year.

Other than hyperscalers, few companies have the combination of demand for heavy AI workloads and the required capital to make investing in custom AI chips worthwhile. A rare example is Tesla, which developed the D1 chip for its Dojo supercomputer. A combination of D1 ASICs and NVIDIA A100 GPUs analyze petabytes of data from the cameras on its millions of vehicles to improve and refine the algorithms for Tesla Autopilot. As AI becomes more ubiquitous, more companies in verticals other than technology may also have the need and scale to develop custom ASICs.

Developers, Developers, Developers

NVIDIA’s CUDA software platform is central to its success in AI. CUDA’s proprietary license is often described as a moat for NVIDIA, as performant competitive hardware must still overcome the friction of software changes required to work on competing systems.

Competitors have been investing in shoring-up out-of-the-box support for open-source frameworks like TensorFlow, ONNX, and PyTorch. OpenAI’s Triton framework supports NVIDIA and AMD GPUs, and Microsoft leverages it for its Maia accelerator. AMD is investing heavily in ROCm and providing more components of that framework as open source, and Intel is investing in oneAPI, which is already fully open source. These frameworks are approaching feature parity with NVIDIA’s CUDA framework.

While Intel and AMD have long-established developer programs and the resources to develop strong software suites, start-ups tend to focus limited resources on hardware —rather than software—development. CUDA’s moat keeps startups at bay, and while those companies can leverage open-source solutions, these do not provide a point of differentiation.

Grow a Spine

One of the most important strategic decisions NVIDIA has made in recent years is acquiring datacenter networking company Mellanox in 2020 for $6.9 billion. Mellanox was the largest provider of InfiniBand networking equipment, which is a popular method of connecting individual compute nodes together for high-performance compute (HPC) systems. NVIDIA uses InfiniBand extensively in its DGX computing appliances, providing a 400 Gbps connection between nodes.

This is where NVIDIA’s strength lies, principally—training larger AI models requires larger-scale systems, which requires high-performance networking to connect individual nodes in compute clusters. While InfiniBand is an open standard, NVIDIA is the only vendor of record for InfiniBand equipment relevant for AI and—like NVIDIA’s GPUs—it comes at a price premium compared to the industry standard: in this case, Ethernet.

The Ultra Ethernet Consortium—which includes steering members AMD, Intel, Meta, Microsoft, Oracle, HPE, Broadcom, Cisco, Arista, and Eviden—aims to iterate on existing Ethernet interfaces to bring Ethernet to 800 Gbps to accommodate the needs of AI workloads and the application processors that run them.

While InfiniBand and Ethernet battle it out for system-to-system interconnect, a similar skirmish is going on between GPU-to-GPU connectivity. NVLink is NVIDIA’s proprietary communications protocol. AMD is spearheading the creation of UALink as an open standard, in cooperation with Intel and Broadcom among others, for GPU-to-GPU connections. Version 1.0 of the UALink specification is expected in Q3 2024, with 1.1 to follow in the next quarter.

Going from specification on paper to product in hand requires a considerable amount of time, likely providing NVIDIA an 18- to 24-month lead to capture more of the market as well as time to iterate and innovate on its existing designs. NVIDIA’s absence among the sea of logos on the websites of these emerging industry consortia—and the noteworthiness of cooperation between Intel and AMD—increasingly gives the appearance of the rest of the semiconductor industry versus NVIDIA.

The World versus NVIDIA?

While it is undoubtedly overstating things to say that the rest of the semiconductor industry is aligned against NVIDIA, it’s certainly true that hyperscalers and many others would prefer to see a more diverse range of suppliers for AI. This is why those companies are diversifying away from NVIDIA, although not so much to Intel and AMD, but more to their own custom silicon. NVIDIA is far ahead of its direct competitors, so the only route available for performance gains are ASICs tailored to specific workflows.

NVIDIA also has a lead in software, although Intel and AMD are catching up. This is where start-ups are typically weak. AMD and Intel have the scale and developer relationships to potentially overcome NVIDIA’s lead more easily than a start-up ever could. However, Intel and AMD’s focus on more open standards may help pave the way for start-ups, which can also leverage those efforts in a way that they cannot leverage CUDA.

Likewise, NVIDIA’s proprietary NVLink protocol and market control of InfiniBand give the company a competitive advantage today. But as more open standards appear, in the longer term these proprietary technologies may prove to be an Achilles heel. Hyperscalers want to be able to mix and match hardware; proprietary tech is only suffered for as long as it provides a significant advantage. NVIDIA has that today, but the next generation of Intel and AMD AI accelerators are due in 2025, along with UALink hardware. Without a surprise technology leap, it’s unlikely that its competitors will overtake NVIDIA next year, but they might at least narrow the gap and erode its dominant market share. 2025 may be the year that NVIDIA’s crown starts to slip, slightly.

Company Profile: IBM

For the past century, IBM has played a prominent role in the research, development, and commercialization of computers. IBM’s heritage as a hardware vendor includes large-scale computers for scientific computing, mainframes for business transaction processing, and microcomputers for personal use—modern PCs extend the architecture of the IBM 5150 Personal Computer from 1981.

Today’s IBM is quite different, as the company has acutely pivoted from its hardware origins in favor of software and services. Since the 1990s, IBM has transitioned out of the commodity hardware business, retaining high-performance systems including the IBM Z and Power series, continuing research into advanced semiconductor process nodes, and expanding research into quantum computing.


Founded in 1911 as the Computing-Tabulating-Recording Company, the “International Business Machines” name was adopted in June 1924. IBM’s first commercial computer was the IBM 701, introduced in 1952, for scientific computing; the IBM 702 was tailored for business transaction processing. In 1959, the first high-volume transistorized computer, the IBM 1401, was introduced. IBM’s S/360 mainframe family, introduced in 1964, is one of the company’s most-enduring designs, with application-level compatibility maintained on present-day IBM Z mainframes.

Multiple IBM inventions are still in use today or are otherwise foundational to modern computing. IBM released the first magnetic hard disk drive in 1956, the first one-transistor DRAM cells in 1966, the floppy disk in 1971, the reduced instruction set computing (RISC) architecture in 1980, and silicon germanium transistors in 1989.

IBM struggled in the early 1990s, bringing on Louis Gerstner as CEO to restore the company. Gerstner was the first “outsider” CEO of IBM in 60 years. IBM shed low-margin component and peripheral businesses, including DRAM (to Toshiba), hard drives (to Hitachi), and printers (forming Lexmark). By the end of the decade, IBM’s prestige was restored through a combination of public relations and renewed credentials in supercomputing, with Deep Blue beating chess champion Garry Kasparov in 1997 and debuting the high-efficiency Blue Gene supercomputer in 1999.

IBM continued divesting low-margin businesses while advancing research work throughout the 2000s: Lenovo acquired IBM’s PC business in 2004 and commodity server business in 2014. Toshiba acquired its point-of-sales hardware in 2012, and GlobalFoundries acquired its semiconductor manufacturing business in 2014.

IBM’s Watson supercomputer beat Ken Jennings and Brad Rutter on the game show Jeopardy! in 2011, echoing the public relations win the company had in 1997 with Kasparov. Watson was subsequently productized for the healthcare market and used in weather forecasting; these pursuits failed to gain traction, with IBM exiting and divesting these ventures in 2022 and 2023, respectively.

Under Gerstner’s leadership, IBM slowly transitioned to a software company, with the purchase of Lotus and Tivoli in the mid-1990s. This continued after Gerstner’s retirement with the acquisition of Rational Software, Cognos, and SPSS in the 2000s. The landmark purchase of Red Hat in 2019 for $34 billion was IBM’s largest acquisition to date. Completing this transition, IBM announced plans to spin out its infrastructure services business as Kyndryl, completing the separation in 2021.

Combined, “IBM Hat” is a primarily software-first company, focusing on hybrid cloud and AI, while retaining its high-value, high-margin mainframe and infrastructure businesses. IBM continues to advance the science of computing with IBM Research, which has specializations in quantum computing and development of leading-edge process nodes—most recently partnering with Japanese startup Rapidus to commercialize IBM’s 2nm technology.

IBM in Numbers

Closing out 2023, IBM recorded $61.9 billion in revenue, up 2% year-on-year, with a gross profit of $34.3 billion, up 5% year-on-year, and net profit of $7.5 billion, at parity with 2022. Owing to its pivot to software—and generally, the predictability of Software as a Service (SaaS) revenue, versus the cyclical nature of hardware—IBM’s software revenues totaled $26.3 billion in 2023, and consulting totaled $20 billion.

Most of IBM’s revenues are linked to its software and consulting lines of business; the former bringing in $5.9 billion in revenue in Q1 2024, the latter bringing in $5.2 billion. Naturally, these operations complement IBM’s infrastructure business, with transaction processing software subscriptions for IBM Z covered in the software unit. The consulting unit leverages partnerships with other software and cloud infrastructure vendors—most notably, Microsoft, AWS, and SAP—for architecting and modernizing an enterprise IT landscape.

IBM’s infrastructure unit brought in $3.1 billion in Q1 2024, representing a 0.2% increase at constant currency. IBM divides its infrastructure unit into hybrid infrastructure—including IBM Z and Power systems—which was up 6%, and infrastructure support which declined 7%. Considering that the current generation of IBM Z debuted in 2022 and IBM Power debuted in 2021, continued growth for hardware this far into the product life cycle is noteworthy.

IBM spent $6.78 billion in research and development (R&D) in 2023, representing 11% of its total revenue for the year. IBM Research’s operations include quantum computing, semiconductor design and packaging, and AI.


IBM Z is the technical successor of the S/360 family of mainframes. The Z represents “zero downtime,” as the servers are architected for redundancy, supporting hot failovers to ensure continuous operation with high sustained CPU utilization and multiple levels of virtualization supported. Enterprise adoption of IBM Z is widespread: it is estimated that 70% of the world’s transactions by value go through an IBM mainframe.

The most recent—IBM z16—was introduced in April 2022, based around IBM’s Telum CPU, which added support for AI acceleration. The z15 mainframe used the z15 CPU; despite decoupling the CPU and mainframe names, Telum is not available outside of z16. IBM CFO Jim Kavanaugh indicated that IBM is working with over 100 clients on the application of AI on z16. Development of IBM Z continues, with details of a future Z system CPU to be announced at the IEEE Hot Chips conference in August 2024.


IBM Power is an alternative instruction set architecture (ISA) that competes against x86 and Arm—unlike IBM’s Z architecture, Power CPUs are available in systems not manufactured by IBM. Modern Power systems are known for scalability and performance for enterprise workloads, such as SAP and Oracle.

IBM Power debuted in the RS/6000 series of servers and workstations in 1990. In collaboration with Apple and Motorola, IBM Power was briefly adapted for consumer electronics as PowerPC, notably found in Apple’s Mac lineup from 1994 to 2005, as well as the Nintendo Wii, Sony PlayStation 3, and Microsoft XBOX 360. In 2008, IBM discontinued the PowerPC line, and rebranded its server offerings as IBM Power.

In 2013, IBM pursued a transformation effort for the Power ISA, aiming for interoperability (e.g., proper little-endian support enabling easy porting of x86-64 applications) and fostering an ecosystem for the architecture. IBM co-founded the OpenPOWER Foundation, which included Red Hat, Mellanox, and NVIDIA. The move also placed IBM in the merchant silicon market as it sold Power CPUs to server OEMs. In 2019, IBM released the Power ISA as open source, allowing third parties to implement the ISA without paying royalties, inclusive of patents.

IBM’s Power9 CPUs were used in the Summit and Sierra supercomputers, alongside NVIDIA Tesla V100 GPUs. Summit and Sierra were the first and second-ranked supercomputers in the TOP500 list when they were first submitted in November 2018; as of June 2024, they are ninth and twelfth, respectively.

While the OpenPOWER Foundation fosters a collaborative development model, IBM remains the driver of Power ISA development and designer of Power CPUs. In 2020, the Power10 CPU was introduced, including prefixed instructions, advanced memory managed, nested virtualization, and SIMD support for AI/ML.

Development of the Power ISA continues, with the first patches for the Linux kernel that will enable the next generation of Power processor-based servers.

IBM & Rapidus collaborate on 2nm node, advanced packaging

In 2021, IBM Research announced the manufacture of the first test chip for its 2nm gate-all-around (GAA) process node. IBM’s research into the method had taken a decade to reach this milestone. IBM Research purchased their first HVM EUV tool from ASML in 2014-15 and were the first research organization to propose introducing EUV into the front end of the line (FEOL). Technical refinements were necessary to realize the GAA device, the first of which is the realization of a dry inner spacer module in 2019, followed by the industry’s first bottom dielectric isolation.

The route to commercialization of IBM’s 2nm node was unclear when the test chip launched. IBM and Samsung collaborated on development of 3nm GAA, and Samsung manufactures IBM’s Telum and Power 10 processors. IBM and Intel announced an R&D collaboration agreement months ahead of IBM’s 2nm announcement in support of what is now Intel Foundry. Of note, in 2018, GlobalFoundries abandoned leading-edge node manufacturing, making IBM’s former fabrication plant an unsuitable receiver of the technology.

In August 2022, Rapidus was founded to bring leading-edge manufacturing back to Japan with a ¥7.3 billion ($55.3 million) investment provided by a consortium led by Denso, Kioxia, MUFG Bank, NEC, NTT, SoftBank, Sony, and Toyota. In December 2022, Rapidus and IBM announced a joint development agreement to manufacture semiconductors on IBM’s 2nm GAA-FET process, with Rapidus estimating the start of mass production in “the latter half of the 2020s,” representing a blisteringly fast timeline for a startup. Rapidus has subsequently received nearly ¥1 trillion in direct subsidies from the Japanese government.


IBM Research has held a decades-long interest in quantum computing, including early quantum information theory work in the 1970s and the demonstration of its first proof-of-concept quantum computer in 1998. IBM’s Quantum Experience cloud platform launched in 2016 with a five-qubit superconducting quantum processor (“Canary”), the first commercially available gate-based quantum computer.

IBM continues to iterate on its gate-based, superconducting qubit architecture. In December 2023, IBM announced Condor, a 1,121-qubit system, as well as Heron, a 133-qubit tunable-coupler quantum processor, which the company touts as eliminating crosstalk errors between qubits, significantly impacting the coherence time in which an algorithm can be usefully executed.


IBM’s role in the overall technology industry has shifted as the enterprise technology market itself has moved, though the company’s status as a research center remains secure amid these changes. Notably, IBM has the distinction of being able to demand attention from mass media for its research activities—for a (now) wholly business-to-business brand, IBM maintains a high profile among consumers.

This research—as with any research—does not always translate into a viable commercial product. Despite IBM’s high-profile Watson win on Jeopardy!, the underlying technology failed to gain traction among enterprises. IBM’s ambition for AI continues with the technologically unrelated watsonx products.

IBM’s institutional willingness to see through research on a long-term basis is worth celebrating. IBM’s decades-long ambitions into quantum computing and continued development of advanced semiconductor manufacturing nodes—despite not operating a manufacturing fab—are contributing to a healthier technology industry.

Interesting Observations

Navigating the Semiconductor Landscape

Delve into market trends, regional dynamics, and future projections in the semiconductor industry. Discover insights from industry leaders and analysts on key topics like lithography advancements, the rise of Chinese manufacturers, and global investment strategies.

Logic Scaling for the Next Decade

This is the first of seven modules that TechInsights is publishing on logic scaling challenges. Starting at the beginning, we introduce Moore’s Law and the Golden Age of scaling followed by a brief history of how scaling was achieved in the past. From reduction of physical dimensions to novel materials introduction, all good things must eventually come to an end, and so it was with planar transistor technology. The transition of planar, 2D transistors to FinFET transistors heralded the next major technology leap to 3D transistor structures and the need for closer collaboration between design optimization and process technology optimization, called design technology co-optimization, or DTCO.

Generative AI in the Telecom Industry

Explore how generative AI is revolutionizing the telecom industry by automating tasks, optimizing networks, and enhancing customer experiences.

Retro Tech

This smartphone—released 15 years ago—was the first competitive answer to the iPhone, combining a modern smartphone touchscreen interface with a slide-out physical keyboard. Supporting iTunes sync out-of-the-box, it angered Steve Jobs and Tim Cook, partially as many of the designers involved were former Apple engineers. Despite initial critical acclaim, business disputes with mobile carrier partners and a famously ill-fated acquisition doomed this phone and the storied brand behind it to the history books.

What was it? Click the image to find out.

Retro Tech

1 thought on “”

Leave a Comment