Newsletter · June 7, 2026
Weekly Digest 23
SpaceX turns its GPU buildout into a third-party product, signing Google to $920M a month for ~110,000 GPUs; Huawei pitches a τ Scaling roadmap to route around EUV; Microsoft ships Scout, an always-on enterprise agent; and an AI-driven memory squeeze spills into chipflation across the hardware economy.
Topics we are tracking
SpaceX's journey to become a cloud provider
Source: Reuters — SpaceX signs cloud deal with Google
SpaceX signed another AI infrastructure deal this week. Google will pay the company $920M per month from October 2026 through June 2029 for access to roughly 110,000 Nvidia GPUs, plus related CPU, memory, and infrastructure capacity. The agreement follows a similar deal with Anthropic, which secured access to SpaceX's Colossus 1 data-center capacity at $1.25B per month.
That makes SpaceX a more serious AI infrastructure company than the market may have expected. Starlink already made it a global communications network. xAI pushed it into large-scale GPU buildouts. The Google and Anthropic contracts now turn that capacity into a third-party revenue product. SpaceX is not becoming AWS, but it is becoming a compute-capacity provider.
The AI cloud market is broadening beyond the traditional hyperscalers because everyone is supply constrained. If a company can assemble power, chips, cooling, networking, and capital at speed, it can rent compute into a market that still does not have enough of it. SpaceX has capital-market access, engineering velocity, satellite connectivity, launch capability, and a willingness to build physical infrastructure unusually fast.
There is still a durability question. Google described the agreement as bridge capacity, not a permanent outsourcing of its AI infrastructure strategy. The deal also includes delivery and termination provisions, so SpaceX still has to prove it can deliver the capacity on time.
There is a second tension inside SpaceX itself. The company is monetising compute externally, but it may also need more of that compute internally. Cursor is already partnering with SpaceX on model training, and SpaceX has an option to acquire the company later this year. If Cursor's Composer models keep improving, SpaceX may want more capacity for both training and inference rather than renting it out to Google and Anthropic.
That is what makes the signal important. Even Google is willing to rent AI compute from a rocket company. In the AI era, cloud infrastructure is being repriced around whoever can assemble chips, power, and capital fastest.
Huawei looks for a way around EUV
Source: Huawei — τ Scaling Law and the LogicFolding architecture (IEEE ISCAS)
Huawei used IEEE ISCAS in Shanghai to present what it calls the τ Scaling Law, a new semiconductor roadmap built around reducing signal delay rather than only shrinking transistor geometry. The company says the approach combines device, circuit, chip, and system-level optimisation, with LogicFolding as the core architecture.
Moore's Law has not stopped, but it has become harder and more expensive. The industry is getting less benefit from simple geometric scaling, while cost-per-transistor improvements have slowed. Huawei is trying to find performance gains outside the classic shrink-the-node playbook.
The company says it has already designed and mass-produced chips based on τ Scaling. The first Kirin chips using LogicFolding are expected to launch in autumn 2026. By 2031, Huawei says its high-end chips could reach transistor density equivalent to a 14Å, or 1.4nm, process.
This is not the same as producing true 1.4nm chips with leading-edge EUV. China still does not have access to ASML's most advanced EUV tools. LogicFolding is an attempt to work around that constraint by shortening critical wiring paths, reducing signal delay, stacking functions more tightly, and optimising the full system rather than only the transistor.
There is still a large gap between a roadmap and a competitive AI chip. The first real test will be the upcoming Kirin chip, where Huawei has claimed meaningful gains in power efficiency and peak speed. Analysts remain cautious because there is not enough independent data on yield, cost, thermal behaviour, or performance against chips built on more advanced nodes.
China is being pushed away from pure lithography scaling and toward architecture, packaging, EDA, and system-level workarounds. Some of those efforts will fail. Some may become durable. EUV export controls remain one of the main hardware constraints on China's AI acceleration. If Huawei can find scalable alternatives, even partial ones, the AI race between China and the US becomes less dependent on one chokepoint.
Microsoft's always-on agent problem
Source: Microsoft — Introducing Microsoft Scout, your always-on personal agent
At Build this year, Microsoft introduced Scout, its first always-on agent for Microsoft 365. Scout is designed to work across Teams, Outlook, OneDrive, SharePoint, email, calendar, contacts, browser activity, local resources, and enterprise systems. It can schedule meetings, prepare materials, track deliverables, block calendar time, and flag stalled decisions.
This is a different product category from the chatbot. Scout is not waiting for a prompt. It sits inside the flow of work, builds context over time, and acts in the background under a governed enterprise identity. Microsoft calls this class of software an Autopilot: an agent that can keep work moving even when the user is not actively interacting with it.
That makes the product useful, but also sensitive. An always-on agent has access to the most important surface in the enterprise: inbox, calendar, documents, chats, meetings, contacts, permissions, and organisational context. It can only be valuable if it is embedded deeply enough to understand how work actually happens.
That creates a large commercial opportunity for Microsoft. The company already owns the enterprise distribution surface. If agents become the new interface to work, Microsoft has the cleanest path to make them default inside large organisations. Scout can sit on top of Microsoft 365, use Work IQ as context, and operate inside existing identity, security, and compliance systems.
This is where enterprise AI is heading. Not more chat windows. Not another sidebar. A persistent agent with memory, identity, permissions, and enough context to act before being asked.
AI is eating the memory market
Source: Reuters — SK Hynix plans to double wafer capacity over the next five years, chairman says
Source: Reuters — Automakers, retailers warn memory-chip shortage is impacting prices
SK Hynix plans to double its wafer capacity over the next five years. The company is already one of the main winners of the AI infrastructure cycle, with a dominant position in high-bandwidth memory, the memory used inside Nvidia's AI systems. Its chairman also warned that shortages may persist until 2030.
The memory bottleneck started inside the data center. It is now moving into the rest of the economy. U.S. industry groups representing automakers, retailers, electronics companies, and telecom firms warned that AI data-center demand is pushing up memory-chip prices and threatening supply availability for non-AI products.
This is how AI capex turns into inflation pressure. Cars, smartphones, PCs, medical devices, telecom equipment, and industrial electronics all need memory. If suppliers prioritise HBM and data-center customers, everyone else competes for what is left. Morgan Stanley has started calling this "chipflation." AI does not only raise the cost of training frontier models. It changes the price of shared inputs across the hardware economy.
The winners are obvious: SK Hynix, Samsung, Micron, and the equipment and materials companies tied to memory expansion. The losers are less obvious but more numerous: consumer-electronics companies with thin margins, automakers still recovering from the last chip shortage, and any hardware company that assumed memory would stay cheap and abundant.
SK Hynix doubling capacity sounds like relief. It may also be confirmation that the shortage is structural. If the leading HBM supplier needs five years to double output and still expects tightness through the end of the decade, memory is no longer a cyclical afterthought; it is becoming one of the core constraints of the AI economy.
The AI memory market: HBM revenue, by year
Seen on X
Other interesting stories
- Bloomberg: Perplexity splits AI work between PCs and servers to ease strain
- Financial Times: China's comparative advantage is industrial policy
- Barron's: The U.S. Government Could Take a Stake in OpenAI. What That Means for Tech Stocks.
- Reuters: Uber, Autobrains target Munich robotaxi rollout
- The Information: OpenAI to release internal tool to weaken Nvidia's software advantage
