Diffusion CapitalDiffusion Capital
Back to insights

Newsletter · August 16, 2026

Weekly Digest 33

Tesla prepares to demo a Roadster that briefly flies on cold-gas thrusters, the US drafts a letter telling partners to pick a side between Pax Silica and China’s new WAICO bloc, and Nvidia tests Rubin Ultra with less HBM than it announced.

Topics we are tracking

Tesla Prepares to Demo a Flying Roadster

Source: The Information — Tesla Nears Unveiling Radical New Flying Roadster

There hasn’t been much excitement around Tesla in 2026 so far. Year to date, the stock is down around 27%. From the outside it sometimes even looks like Elon is losing interest in Tesla and focusing more of his attention on SpaceX. SpaceX completed its $60 billion acquisition of Cursor this week and is increasingly competing with OpenAI and Anthropic at the frontier of AI, an area that seems to have captured more of his attention recently.

The Information reported this week that Tesla is preparing to demo two versions of the long-delayed Roadster at the SpaceX test facility. One will be a regular version of the Roadster, while the other will be able to briefly fly using cold-gas thrusters developed with SpaceX.

The second-generation Roadster as Tesla showed it in 2017. Neither of the two cars being prepared for the demo has been pictured.

Musk has been talking about this for a long time. Tesla first unveiled the second-generation Roadster in 2017, and Musk announced the “SpaceX option package” a year later, saying it would use around ten small rocket thrusters to improve acceleration. During his latest appearance on Joe Rogan, Musk also suggested that the Roadster could actually fly.

According to The Information, the demo will include driving a remote-operated car on a magnetized roller-coaster-like ramp, including driving upside down, as well as hovering in the air. All participants in the live event also need to be stationed several hundred yards away because of potential damage to their eardrums. It’s perhaps not surprising that this car will not be legal to drive on any streets, will only be produced in very small quantities and will cost up to a million dollars. While this all sounds a bit ridiculous, time has shown that dismissing Elon’s more ambitious engineering projects is often a mistake.

The project itself raises more questions than answers right now, but it’s another datapoint that Elon can point to where SpaceX and Tesla overlap and where a merger would make sense. The two companies are already collaborating on batteries, Starlink and now semiconductor manufacturing with the newly announced TerraFab plant. He already mentioned the growing overlap between the two companies in the most recent earnings call, vaguely answering an analyst’s question about a potential merger.

The US and China Build Competing AI Blocs

Source: Reuters — US to tell partners they must pick sides in AI race with China

The US is preparing to tell countries that they may have to pick a side in the AI race with China. According to Reuters, the State Department has drafted a letter warning countries that they could be excluded from the US-led Pax Silica initiative if they also participate in competing Chinese AI initiatives.

The Pax Silica initiative was launched by the US last year to better coordinate the supply chains needed for AI. It covers models, semiconductors, critical minerals and energy, and members are expected to work together on investment and export controls. Many countries around the world have joined so far, including Japan, South Korea, Australia and India. The EU also joined in June, together with several countries in Latin America and Kazakhstan.

This latest escalation is mainly about Kazakhstan. It is interesting to the US because of its large critical mineral reserves. Last year the country announced it had discovered a 20-million-ton rare earth metals deposit. This would give Kazakhstan the third-largest rare earth deposit, behind only China and Brazil. A week after signing on to the Pax Silica, the country also became a founding member of China’s World Artificial Intelligence Cooperation Organization, potentially threatening Western access to this repository and reducing the chokehold China has on rare earth metals.

Exhibit 1. Two blocs, one pivot

KazakhstanIn both, one week apart
Pax Silica (45)WAICO (30)Member of both (1)Not alignedSmall states shown as markers
Membership as of 16 August 2026. The EU signed the Pax Silica declaration as a bloc on 23 June, so all 27 member states are shaded — the Netherlands, Germany, Greece, Sweden, Finland and Estonia had also signed individually. WAICO counts its 29 founding signatories from the Shanghai agreement of 16 July plus Iran, which acceded afterwards. Kazakhstan is the only country in both. Sources: US State Department, ITIF, Reuters, Xinhua, CGTN.

China launched WAICO in July as an alternative framework for international AI cooperation. The organization will be headquartered in Shanghai and is focused on AI development and governance. China has also been offering developing countries access to training, AI infrastructure and Chinese technology. This looks similar in theory to China’s Belt and Road Initiative, and some scholars have already drawn that comparison. However, it is not entirely clear how much AI infrastructure China can provide to third countries given that it has about 13–15x less advanced compute available than the US.

China has been very successful in developing frontier models such as GLM and Kimi, but exporting the models without the compute infrastructure will only get countries so far. The broader Chinese offer could still be attractive when combined with infrastructure investment and financing. For now, China’s ability to export AI compute is constrained by its domestic semiconductor capacity. Until China can produce advanced logic and memory chips at a much higher rate, it will remain limited in how much compute infrastructure it can offer to its partners. The main bottleneck remains access to advanced lithography equipment, particularly EUV and advanced immersion DUV systems.

The Pax Silica initiative covers large parts of the AI supply chain, and it would be in the interest of every signatory to collaborate more closely to compete with China’s manufacturing and industrial capacity as well as its dominance in critical minerals. The problem is that the US under President Trump has increasingly adopted a more mercantilist approach to foreign policy, which makes this kind of cooperation more difficult. Building an alternative to the Chinese supply chain will require the US, Europe and Asian allies to invest together and rely on each other’s strengths rather than impose tariffs on each other.

Nvidia testing lower memory configs for Rubin Ultra

Source: The Information — Nvidia Weighs Radical Idea: Less Rubin Ultra Chip Memory

Nvidia is reportedly testing versions of Rubin Ultra with less HBM than originally planned. The Information reported that Nvidia is testing prototypes with 192 GB and 256 GB of memory. The original configurations for Rubin Ultra were much larger. The company is also considering using HBM4 instead of the announced HBM4E.

The standard Rubin comes with 288 GB of HBM4 and is already in production. Micron, SK hynix and Samsung have all started shipping HBM4 in 2026. HBM4 introduced a more advanced, customized logic base die, which allows for closer integration with the accelerator. Both Micron and SK hynix are working with TSMC on this, while Samsung is using its own foundry.

HBM4E is supposed to start production in 2027, and its design spec pushes this custom logic further, with TSMC discussing moving it to N3. Rubin Ultra was also expected to use 16-high HBM4E stacks, compared with the 12-high stacks used for standard Rubin. This adds additional complexity around bonding, thermals and yield.

This might also tie into the reports that Nvidia cancelled its quad-die Rubin Ultra design and is moving to a dual-die design due to issues with the manufacturing complexity. Nvidia has yet to confirm this, but if true, the move to less memory would make more sense. Moving to two compute dies would significantly reduce the amount of compute in each package and therefore also reduce the amount of memory capacity and bandwidth needed to keep those dies utilized.

Using less HBM per GPU could allow Nvidia to ship more GPUs with the available memory. It’s clear that the memory shortage is starting to affect Nvidia’s product roadmap. SK hynix CEO Kwak Noh-jung recently said that 2027 will be the worst year for the memory shortage, and that supply constraints will last through 2030. SK hynix is ramping up capacity, but new fabs take years to build. If Nvidia expects memory to remain constrained for several more product generations, it makes sense to redesign its product around the amount of HBM it can realistically source.

Seen on X

Other interesting stories