Starship Orbital, TPUs in Space, Grok 4.8? | ep 22

Insights, discussion points, and conclusions

1. Starship’s orbital milestone shifts attention toward recovery and reuse

Discussion · 02:18–05:10. The speakers describe the change of plan during the flight, when SpaceX proceeded with the final burn to reach orbit. Aaron argues that earlier engine relights and demonstrated flight performance had already answered much of the technical question about reaching orbit. He describes the remaining maneuver as a roughly 19-second burn on one engine. Both speakers see the booster and ship’s controlled descents as encouraging signs for future catches.

Conclusion. Their assessment is that recovery and repeatable reuse now deserve more attention in judging Starship’s progress. A controlled descent supports the case for an attempted catch, but this flight did not establish routine recovery or rapid reuse.

2. Heat-shield performance and engine shutdowns remain open engineering questions

Discussion · 05:10–09:13. A viewer asks about the heat shield because the broadcast provided limited close-up footage after landing. The speakers report seeing little visible streaking but have no additional inspection evidence. They also discuss engine shutdowns and a reported issue with ice ingestion. Aaron suggests that plumbing and integration changes could be involved, while explicitly acknowledging that he lacks inside engineering information. He notes the vehicle’s ability to continue after an engine shuts down and the limited flight history of the V3 design.

Conclusion. They remain positive about the program’s direction, while leaving the precise causes of shutdowns and the heat shield’s condition unresolved. Further flights and recovered hardware should provide better evidence. The discussion offers an interpretation of the available footage rather than a verified engineering diagnosis.

3. New Glenn’s proposed launch cadence would matter if Blue Origin achieves it

Discussion · 11:28–14:32. Vlad describes an assessment submitted to the Federal Aviation Administration seeking an increase from 12 to 50 New Glenn launches per year. He also mentions NASA adding another New Glenn variant to its launch contract. The speakers view those developments as signs of Blue Origin’s ambition, while questioning when the company could reach that cadence.

Conclusion. Fifty annual New Glenn launches would materially increase launch supply, but permission to support that cadence would still need to be followed by operational performance. No firm timeline is established. Aaron briefly discusses a model result of roughly 40 launches in 2027, then clarifies that he was referring to Starship; that number should not be read as a New Glenn forecast.

4. Google’s TPU test gives orbital computing a larger technical sponsor

Discussion · 14:32–18:30. The speakers discuss a planned Falcon 9 launch later that day carrying Google TPUs on Planet spacecraft. They compare it with Starcloud’s earlier deployment of an NVIDIA H100 AI accelerator. Aaron argues that Google’s involvement gives the field a sponsor with substantial computing expertise and capital, along with leadership support for Project Suncatcher, its orbital computing effort.

Conclusion. They consider the mission an important experiment for learning how AI chips operate in space. The broadcast takes place before the planned launch and does not report its outcome. Their enthusiasm concerns the sponsor and the testing opportunity; a prototype alone does not establish commercial data-center economics.

5. Distributed orbital computing must account for the cost of connecting its chips

Discussion · 18:30–28:20. The speakers compare a formation of small satellites separated by roughly 150 meters with a larger spacecraft carrying more computing hardware. When a large AI model runs across multiple chips, those chips must exchange data fast enough to keep the workload moving. Mach33’s earlier orbital containment analysis examines how the choice to concentrate or distribute hardware changes spacecraft overhead and communication requirements.

Vlad argues that a formation duplicates spacecraft structures and other supporting hardware across more vehicles. A larger spacecraft can reduce some of that duplication, although it introduces its own power, cooling, and manufacturing challenges. Aaron raises a hybrid arrangement with fewer, larger spacecraft that still fly near one another.

Conclusion. The speakers see a possible economic advantage in concentrating more computing hardware on each spacecraft. They leave the eventual architecture open because its practicality depends on spacecraft production, launch access, and the requirements of the computing workload.

6. Available spacecraft and launch vehicles constrain Google’s initial design

Discussion · 20:30–28:20. Aaron suggests that Planet’s experience making smaller satellites helps explain Google’s initial approach. Vlad adds that Google must plan around available launch services and established spacecraft suppliers. He expects SpaceX to prioritize its own satellite deployments when allocating Starship capacity, which would complicate planning around large spacecraft that require that vehicle.

Conclusion. They interpret Suncatcher’s formation-flying design partly as a response to current supplier and launch constraints. An eventual move toward larger spacecraft is a possibility to watch as the partners gain flight data and manufacturing experience. The speakers present this explanation as their inference, with no claim to know Google’s internal decision process.

7. Radiation, cooling, and formation control require flight evidence

Discussion · 22:02–28:20. Aaron identifies radiation behavior and thermal performance as useful results to watch from the TPU test. He expresses confidence that radiation will be manageable, but the discussion provides no mission results. Vlad argues that nearby satellites sharing an orbit should be understood through their relative motion rather than by comparing them with vehicles moving through traffic on Earth. The speakers also lack exact dimensions and radiator specifications for the test spacecraft.

Conclusion. The mission should help resolve hardware questions that remain open in the broadcast. Their confidence in formation flying and chip survivability remains an expectation to test. The conversation does not establish collision probabilities, required cooling area, or radiation tolerance.

8. xAI’s compute arrangement with Anthropic receives only a brief assessment

Discussion · 28:20–28:45. The speakers mention supplementary computing capacity that xAI is supplying to Anthropic. They treat it as a favorable revenue development and move on to the Grok analysis.

Conclusion. Their takeaway is that selling compute can generate revenue for xAI. They do not assess the agreement’s terms, margins, or effect on xAI’s own model-training schedule.

9. The revised Grok forecast corrects an assumption about the software rewrite

Discussion · 28:41–37:12. Vlad says the earlier forecast assumed that both pretraining and post-training would move to xAI’s in-house C++ software. Pretraining builds a model’s initial capabilities from data; post-training adjusts its behavior afterward and can include reinforcement learning. The speakers say later comments from Elon Musk indicated that reinforcement learning remained on the older third-party software, explaining why their previous estimate for Grok 4.7 was too early.

Aaron explains how software overhead can matter when a training process repeatedly generates model responses. He also notes that replacing reinforcement-learning software involves multiple workflows and requires more work than the team had assumed.

Conclusion. They revise the forecast to distinguish the completed pretraining rewrite from the unfinished reinforcement-learning rewrite. Their working assumption is that Grok 4.8 will still use the older reinforcement-learning software. Its completion date remains unknown.

10. Longer model responses can outweigh some gains from additional computing capacity

Discussion · 37:39–44:46; 49:32–50:13. The speakers discuss a reported change in Grok 4.8 training toward longer responses, after an earlier penalty for long answers that they say caused Grok 4.7 to abandon difficult tasks too soon. Generating more tokens, the units of text a model produces, lengthens each response used during reinforcement learning. Their model also assumes a modest increase in model size and more available computing capacity.

Vlad presents an estimated 5.3 weeks of reinforcement learning for Grok 4.8, compared with 4.3 weeks for Grok 4.7. Under the older software, the model estimates diminishing scheduling benefits beyond roughly 437,000 H100-equivalent chips, even though roughly 831,000 equivalents could be available for reinforcement learning. H100 equivalents express a mixed fleet’s capacity relative to NVIDIA’s H100 chip.

Aaron also distinguishes a reported future goal of allocating about 10% of computing capacity to training from the current allocation. He argues that diminishing gains from more training hardware give xAI a technical reason to allocate additional chips to inference, the process of running a trained model for users. Vlad cites a modeled power requirement of 1.66 gigawatts for the larger fleet discussed in the analysis.

Conclusion. Within their model, longer responses and software limitations carry more weight than adding further chips once that threshold is reached. They expect more incremental capacity to serve users as training gains diminish. The fleet and power figures are estimates discussed in the episode, rather than disclosed operating measurements from xAI.

11. Grok’s release dates provide a test of the model’s assumptions

Discussion · 29:11–30:04; 44:48–50:13; 57:45–58:05. The speakers compare release scenarios that vary response length and reinforcement-learning software. The opening forecast gives a 46–83-day interval and an 18 November central date for Grok 4.8. A later scenario chart gives 6 November–13 December, with 21 November as its median, for the combination of longer responses and the older software. They close with a broader expectation of mid-November for Grok 4.8 and the last week of January for Grok 4.9.

In the sensitivity discussion, retaining shorter responses moves the modeled release roughly 11 days earlier. A hypothetical in-house reinforcement-learning stack produces a modeled 2.2-fold speed improvement and roughly 27 days of schedule savings. Another scenario places the release around 26 October if the new stack is already operating.

Conclusion. Their shared directional forecast is mid-November 2026 for Grok 4.8, followed by late January 2027 for Grok 4.9. The episode contains slightly different central dates and ranges, so it supports a conditional forecast rather than one exact deadline. Release timing would help them update their assumptions, although a date alone would not prove which internal change caused it.

12. Larger frontier models could take longer to develop as the constraint changes

Discussion · 50:20–57:40. Aaron predicts that fast releases of smaller or specialized models could coexist with longer intervals between much larger frontier models. He expects reinforcement learning across many tasks to take longer as model size grows, making software improvements increasingly useful. The speakers then consider data supply and quality as subsequent constraints. They discuss continuously refreshed sources such as X and YouTube, while acknowledging uncertainty about larger models’ actual sizes and data requirements.

They also expect competing labs to pursue similar software improvements. Aaron argues that the ability to rewrite and adjust these systems quickly could matter, but leaves any lasting advantage for xAI unproven.

Conclusion. The speakers propose watching separate release trends for everyday models and larger frontier models over the following six months. Their longer-term thesis is that the limiting factor can move from computing capacity to software throughput and eventually toward suitable training data. This is a forecast, with no measured threshold or demonstrated industry-wide slowdown established in the episode.

Source and editorial note

This recap follows the October 1, 2026 episode, which runs approximately 59 minutes. Timestamps locate the discussions in the recording. Claims about company plans, flight performance, and AI development are attributed to the speakers and have not been independently fact-checked for this recap. Model estimates and predictions remain conditional. The transcript has been summarized for clarity, with unresolved questions and conflicting forecast dates preserved.