Weekly Digest 36

Topics we are tracking

OpenAI Launches Astra

Source: OpenAI — GPT-6 Astra

This week OpenAI launched their long-awaited Astra model, or GPT-6 Astra. Based on the engagement online it seems to be their most successful launch since Sora was first introduced, and OpenAI President Greg Brockman himself called it AGI.

We are not going into detail about specific benchmarks here, but it is noticeable that Astra is beating the competition across a wide variety of benchmarks. Especially the computer use, game building and 3D generation capabilities seem to be a step-function leap forward, with Astra being able to use a variety of computer programs seamlessly.

In their blog post OpenAI called it their most aligned model yet and have shown data to support that claim. However, some researchers have pointed out that the recurrent depth technique used by Astra will make it harder to monitor the chain-of-thought reasoning. The technique allows the model to spend additional computation refining its internal representations through repeated passes, without expressing each intermediate step as another reasoning token. Researchers have previously used this approach in smaller models, showing that additional passes can improve reasoning performance. Google DeepMind and Anthropic are reportedly also considering using it.

There has long been debate over the tradeoff between human-readable chains of thought and efficiency. OpenAI's Chief Scientist Jakub Pachocki has reiterated the lab's commitment to legible chains of thought in a post on X.

Astra's use of this technique is limited, but it further adds to the debate about alignment and how powerful agents should be monitored, especially in the aftermath of the Hugging Face report published by Redwood Research and METR. Chain-of-thought monitorability has many limitations and rests on the assumption that it reflects the model's actual internal reasoning, rather than a "fake" reasoning trace designed to mislead the reader. It is flawed but remains one of the most widely used tools in alignment research—and useful for as long as the assumption of faithfulness holds.

Samsung Regains Share in HBM

Source: Counterpoint Research — Global DRAM and HBM Market Share

SK hynix developed the first HBM chips in collaboration with AMD. The idea was to stack DRAM dies vertically and place them alongside the GPU within the same package, allowing a much wider connection between memory and processor. SK hynix has since established itself as the leading HBM supplier and a key memory partner to Nvidia.

Data published this week by Counterpoint Research estimates Samsung's share of HBM revenue to be at 33% in the second quarter, up from 21% in the previous quarter. SK hynix remains the market leader, but Samsung is narrowing the gap.

A line chart of five quarters from the second quarter of 2025 to the second quarter of 2026. SK hynix falls from 64 per cent to 50 per cent. Samsung rises from 15 per cent to 33 per cent, with the whole gain in the latest quarter. Micron holds around 21 per cent before slipping to 18 per cent.

SK hynixSamsungMicron

Share of global HBM revenue. Values may not add to 100% due to rounding. Source: Counterpoint Research Memory Tracker and Forecast Insights Report, Q2 2026.

Source: Counterpoint Research · 2026-09-04

Samsung has struggled with HBM3E due to design and manufacturing issues. SK hynix has adopted a different technique in stacking DRAM dies which provided better heat dissipation and higher manufacturing throughput.

HBM4 involves substantial technical change and has given Samsung an opportunity to improve their position. They have adopted changes to allow for better power management and higher performance, including moving the logic base die to its own 4nm process. The newly introduced Jalapeno chip from OpenAI is reportedly also using HBM4 from Samsung. Further market-share gains will depend on Samsung's ability to maintain high yields as production expands. TrendForce recently reported that Samsung can deliver yields of up to 80%.

HBM3E still accounts for most industry HBM revenue at the moment, so the recent recovery cannot be cleanly attributed to HBM4. Both SK hynix and Samsung are benefiting from tight supply and rising prices for memory. Contract prices for DRAM are expected to rise 13–18% quarter on quarter, with NAND prices expected to rise 10–15% in the same period.

In Nvidia's most recent earnings they reiterated supply constraints for memory. The company projected 70% revenue growth for 2028 but indicated that demand could support a doubling of revenues.

Elon Is Making Turbine Blades

Source: The Information — Exclusive: SpaceX Lays Groundwork for Turbine Blade Factory to Solve Data Center Power Crunch

When Elon was on the Cheeky Pint podcast earlier this year he alluded to the difficulty of getting enough electricity to AI data centers and identified turbine blades and vanes as the bottleneck. In typical Musk fashion, he proposed bringing production in-house: "I think SpaceX and Tesla will probably have to make the turbine blades, the vanes and blades, internally."

This week The Information reported that SpaceX is preparing to do exactly this. The company is preparing a foundry in Texas to manufacture gas-turbine blades and vanes. Musk has confirmed the plans, saying that bringing casting in-house would accelerate turbine development by up to 18 months.

GE Vernova, Siemens Energy and Mitsubishi Heavy Industries account for more than 70% of global gas-turbine production capacity. GE Vernova reported 116 GW of gas-power equipment backlog and production-slot reservations at the end of the second quarter. It plans to increase its annual turbine production rate from 20 GW in Q3 2026 to 30 GW in 2030, illustrating the gradual expansion of manufacturing capacity despite strong demand. Many companies across the supply chain have experienced previous downturns, leaving them cautious about investing in capacity that could become surplus if demand weakens.

The blades and vanes are among the most difficult-to-manufacture components. Turbine blades are made from nickel-based superalloys designed to retain their strength under extreme heat and mechanical stress. Some are cast as a single crystal, while internal cooling passages and protective coatings help keep the metal within its operating limits.

The coatings also rely on rare earth metals. One widely used heat-resistant material is yttria-stabilised zirconia, which contains yttrium oxide. China produces most of the world's yttrium and requires licences to export it.

While this would be a major announcement for almost any other company, it seems like a reasonable endeavour for one that earlier this year announced a 100-million-square-foot fab and whose latest car is supposedly able to "hover" above the ground.

Seen on X

Other interesting stories

• Bloomberg: China Unlikely to Match ASML's Top Tool in Next Decade, UBS Says
• Mint: China's Nvidia alternative? DeepSeek plans massive Huawei AI chip rollout at Inner Mongolia data centre
• Reuters: Broadcom forecasts quarterly revenue below estimates
• The New York Times: How a Blacklisted Chinese Tech Giant Kept Buying America's Best A.I. Chips
• Sinocities: Introducing China Innovation Atlas
• Republic of Science: Industrial Overcapacity Enables Scientific Discovery