Newsletter · June 28, 2026
Weekly Digest 26
Qualcomm buys Modular for $4 billion to own the software layer that lets AI inference move across any chip; Satya Nadella argues AI must get cheap enough to spread through the economy, judged against labor cost rather than legacy SaaS budgets; The Information finds Nvidia's inference-market share is actually rising as production serving rewards software depth; and Goldman Sachs makes the case for South Korea as the missing supply chain for humanoid robots.
Topics we are tracking
Qualcomm buys Modular for $4 billion
Source: Reuters — Qualcomm to buy AI startup Modular for $4 billion
Qualcomm wants to become more than a handset-chip company. At its investor day, it told investors that data centers could become a $15 billion business by 2029, with revenue already expected to reach $5 billion in fiscal 2027. It also named Microsoft and Meta as early customers and said two additional hyperscalers had signed on for custom chips.
Qualcomm revenue mix: FY2025 → FY2029 target
- Handsets
- IoT + Auto
- Licensing
- Data Center
The company is extending its offering beyond pure hardware and is trying to sell a more efficient compute architecture for inference, built around performance per watt, lower memory cost, and a software layer that can make heterogeneous hardware usable.
Modular builds the software layer between AI workloads and the hardware they run on. Its platform is designed to let models run across CPUs, GPUs, NPUs, and custom accelerators without forcing developers to rewrite code for every processor.
Modular gives Qualcomm a developer-facing layer for a world where AI compute is becoming fragmented across GPUs, CPUs, NPUs, custom ASICs, edge devices, and cloud clusters. In that world, the valuable control point is the layer that lets customers move workloads across chips without rebuilding their infrastructure each time.
Performance portability has disappointed before. Developer ecosystems are hard to acquire. And Qualcomm is entering a market where hyperscalers already build their own chips, Nvidia owns the default software stack, and Broadcom and Marvell are capturing custom-silicon demand.
Qualcomm is betting on a more portable AI compute market, but not all workloads will move in that direction. The largest hyperscalers are pushing deeper into hardware-software co-design, with custom chips, custom compilers, custom networking, and models tuned to their own infrastructure. That will remain the winning model for the most performance-sensitive workloads.
The question is what happens below that layer. Most customers do not want to rebuild their stack every time they change chips. They want lower cost, lower lock-in, and enough performance to run inference reliably. That is the market Modular gives Qualcomm a shot at: fragmented AI workloads that need to run across CPUs, GPUs, NPUs, custom accelerators, edge devices, and cloud clusters.
There is always a price to portability. A workload that runs across many architectures is rarely as efficient as one built for a single stack. Qualcomm's bet is that, for a large part of inference, the value of flexibility will outweigh the loss of perfect optimisation.
Nadella wants cheaper intelligence
Source: WSJ — Microsoft's Satya Nadella: We can't let AI giants eat the economy
Satya Nadella warned this week that a few AI giants cannot be allowed to "eat the economy." His argument is that AI needs to become cheap enough to spread through the economy, rather than sitting as a new tax on top of it. If every useful workflow requires expensive frontier-model calls, the value created by AI flows back to the model providers, chip suppliers, and cloud platforms. The customer gets a better tool, but also a new cost line.
Intelligence per dollar, 2023–2026
That argument is right, but the current debate around AI costs often uses the wrong denominator. Most companies still compare AI spend to historical SaaS budgets. Against that benchmark, AI looks expensive. A $100 or $200 per-seat AI product looks absurd if it is judged against Slack, Notion, or a normal software add-on.
AI should not be treated as another SaaS category. In many workflows it substitutes for, compresses, or augments human labor. The better comparison is not AI spend versus SaaS spend. It is AI spend versus software, salaries, external services, management time, and operational delay. The relevant benchmark is labor cost, not software cost.
That is where Nadella's argument about fungibility matters. Companies need to route tasks to the cheapest adequate system: frontier models for hard reasoning, smaller models for routine work, open-weight models for internal deployments, and customer-specific models for proprietary workflows. The cost curve improves when intelligence becomes more substitutable. But frontier intelligence should still be evaluated against human labor, not legacy SaaS. A powerful model may look expensive as software and cheap as labor.
Nvidia's inference share is rising
Source: The Information — Nvidia's share of the AI inference chip market appears to be rising
The strongest bear case on Nvidia has always been that inference will not look like training.
Training large models is where Nvidia's dominance is easiest to understand. The workloads are massive, centralised, and performance-sensitive. Customers care about time-to-train, cluster reliability, networking, and access to the best accelerators. Inference was supposed to be different. Once models move into production, the workload becomes more cost-sensitive and more fragmented. That should create room for custom ASICs, hyperscaler chips, CPUs, NPUs, and cheaper accelerators. That is the Wall Street narrative. Nvidia owns training, but inference is where the market opens up.
The Information's latest estimate cuts against that view. Nvidia's share of the inference chip market has reportedly increased over the past year. Inference engineering has become one of the most sought-after disciplines in Silicon Valley. Running models in production requires batching, memory management, quantisation, low-latency serving, optimised kernels, model compatibility, networking, monitoring, and predictable reliability. Much of this is abstracted away by inference frameworks such as vLLM, SGLang, and TensorRT. TensorRT has been built by Nvidia and is widely regarded as one of the strongest inference frameworks for large production setups on Nvidia hardware.
Custom ASICs will have a role in inference, especially where the customer controls the model, the serving stack, and the utilisation profile. Google's TPUs have already shown the power of hardware-software co-design. But that path takes time. Google is now many TPU generations into that effort. Most buyers cannot replicate that depth of co-design quickly, and most workloads still need flexibility across models, tooling, and deployment environments.
Over the next two to three years, while accelerator supply remains tight and inference demand keeps expanding, Nvidia is likely to remain the default production path. Even after supply improves, its software expertise should continue to defend its position.
South Korea's robot supply chain
Source: Goldman Sachs — South Korea's growing role in humanoid robot development
Goldman Sachs published an interesting note this week on South Korea's role in humanoid robot development.
South Korea already has a deep automotive and electronics supply chain. Many of the capabilities required for humanoids sit adjacent to capabilities Korean suppliers have spent decades building for cars, factories, displays, batteries, and industrial automation. If humanoids move from prototypes into production, the bottleneck will be whether the physical system can be manufactured at cost, with enough reliability, torque density, battery life, and manipulation capability to be useful.
China is the reason this matters strategically. Goldman estimates that 10,000 to 15,000 humanoids were deployed in China in 2025, compared with only hundreds in the U.S. and Korea. Physical AI does not learn from the internet in the same way language models did. It learns from interaction with the world. Every factory deployment, failed grasp, logistics run, and recovery motion becomes part of the data loop. China therefore has two advantages at once: the manufacturing base to make robots cheaper and the deployment base to generate embodied data faster.
The U.S. has leading AI labs and robotics startups, but not the same dense electromechanical manufacturing base as East Asia. Korea could be the missing link: allied with the U.S., deeply industrial, strong in autos, batteries, electronics, displays, and factory automation, and connected to Boston Dynamics through Hyundai.
The model side of robotics is starting to compound. Vision-language-action models, world models, simulation, synthetic data, reinforcement learning, and teleoperation are improving at the same time. Better models make robots more useful; deployed robots generate trajectories, failures, and edge cases; that data improves the next generation of models.
There are, however, physical bottlenecks starting to emerge. A robot still needs actuators, motors, hands, batteries, sensors, controllers, and safety systems that can work reliably at volume. If the body remains expensive, fragile, or supply-constrained, the model flywheel slows down.
Korea already has much of the industrial base required to scale the physical side of robotics: autos, batteries, electronics, displays, factory automation, and precision components.
