Elon Muskwrote on X at 09:35 UTC on 10 November that xAI has “dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time”. He said Grok 4.8’s pre-training is done, that “4.8 is in RL and the evals are starting to climb”, and that xAI will do “a soft rollout to Bot users later this week”.
The post came 57 days, by our arithmetic, after Musk wrote on 14 September that Grok 4.8 “will finish training this week and start RL”. He was replying to an X user who asked “when to expect Grok 4.8”. Musk gave no benchmark figures and no date beyond “later this week”.
Business Pill · Vertical Integration
An evergreen Business Engineer explainer on vertical integration: owning more of the stack you depend on. It explains the general idea only and does not discuss xAI or Grok.
The key insight: Musk wrote on 10 November that xAI has dumped JAX and SGLang for its own C/C++ code across RL and inference, and that Grok 4.8 is in RL, 57 days after he wrote that it would start RL that week, by our arithmetic.
Musk wrote: “Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right.” SFT is supervised fine-tuning, the stage where a model is trained on curated example answers. RL, reinforcement learning, is the stage after it, where the model is trained against rewards.
He ended with a comparison to the current model: “Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.” That evals are climbing is Musk’s statement; he published no scores, and we have none to report.
Calendar days from each event to Elon Musk’s 10 November 2026 post saying Grok 4.8 is in RL: his 14 September post saying 4.8 would finish training that week and start RL (57 days), xAI’s Grok 4.7 launch on 21 September (50), Google’s Gemini 4 Carbon announcement on 28 October (13) and Anthropic’s Claude Fable 5.2 announcement on 9 November (1). Data: Musk’s posts on X, xAI, Google and Anthropic; our arithmetic.
What Was Posted on X
We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.
Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.
Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.
Elon Musk posted this at 09:35 UTC on 10 November. He wrote that xAI has dumped JAX and SGLang and rewritten its RL and inference stack in C/C++, that pre-training is done, that 4.8 is in RL with evals he says are starting to climb, and that a soft rollout to Bot users will come later this week. The post gives no evaluation scores.
From JAX and SGLang to In-House C/C++
xAI built Grok on outside software from the start. Its November 2023 launch post said: “To create Grok, we built a custom training and inference stack based on Kubernetes, Rust, and JAX.” JAX’s GitHub page describes it as “a Python library for accelerator-oriented array computation and program transformation” and says: “This is a research project, not an official Googleproduct.”
SGLang describes itself as “an open-source inference framework for large language, vision-language, and diffusion models, optimized for agentic workloads, large-scale serving, and RL rollouts”. It is hosted by LMSYS. Inference is the serving step, when a trained model answers users.
On 14 September Musk called Grok 4.8 “a 2.5T model trained with our new C++ software stack”. The 10 November post adds the RL and inference stack to that rewrite. Read together, the two posts describe xAI’s own C/C++ code replacing JAX and SGLang across pre-training, RL and serving.
What the Post Says About Grok 4.7
xAI launched Grok 4.7 on 21 September, 50 days before Musk’s post, by our arithmetic. Its launch page called 4.7 “our most capable model for coding and knowledge work” and said it “uses a new, larger base model compared to Grok 4.6”. It priced 4.7 “starting at $2 per million input tokens and $6 per million output tokens”.
Musk did not say what was wrong with 4.7. His words, “bear no resemblance to 4.7 (thankfully)” and “this time, the foundations are built right”, set 4.8 apart from the model xAI sells today. They follow a week of Grok Bot complaints: on 9 November Musk promised overnight fixes after paid users said their usage was running out. Similarweb estimated the Grok app at 93.4 million monthly users in October, a third-party figure, not one from xAI.
The Compute Behind the Rewrite
xAI now sits inside SpaceX; its website footer reads SpaceX Super Intelligence LLC. SpaceX’s 10-Q shows its AI segment took $16.444 billion of $19.593 billion of third-quarter capital expenditures, 84%, by our arithmetic. Its prospectus says Anthropic agreed to pay $1.25 billion a month for compute that “includes approximately 325,000 NVIDIA GPUs”. Epoch AI estimates that SpaceX’s Colossus 2 in Tennessee is the world’s largest AI data center.
The software stack is what turns that hardware into trained models and served tokens. Musk’s post says xAI now writes that layer itself, and that doing so “took some extra time”.
Timing Against Fable 5.2 and Gemini 4 Carbon
The soft rollout Musk describes for this week follows two rival releases. Anthropic announced Claude Fable 5.2 at 18:57 UTC on 9 November; AWS’s price file listed it at $8 input and $40 output per million tokens, down from $10 and $50 for Fable 5.1. Grok 4.7’s launch page had benchmarked 4.7 against Fable 5.1.
Google Cloud’s public catalog listed Gemini 4 Carbon on 10 November at $4.00 input and $20.00 output per million tokens online; Google announced Carbon on 28 October and has not announced general availability. Musk’s post names Grok Bot users for the 4.8 soft rollout and says nothing about API access or price.
The Structural Read
As we read it, the two posts together describe xAI moving from open-source frameworks it did not write to code it owns, at every stage from pre-training to serving; Musk’s own cost line for that move is “some extra time”.
As we read it, Musk set 4.8 against the model xAI sells today: he called 4.8 something that will “bear no resemblance to 4.7 (thankfully)”, 50 days after 4.7’s launch, by our arithmetic.
FOR XAI’S ENGINEERING Musk says the RL and inference stack was rewritten from scratch in C/C++; on 14 September he said 4.8 was trained on a new C++ stack.
FOR GROK BOT USERS Musk says a soft rollout of 4.8 to Bot users comes later this week; he gave no scope, date or price.
FOR READERS COMPARING MODELS Musk published no Grok 4.8 scores; Anthropic announced Fable 5.2 on 9 November, and Google’s catalog listed Gemini 4 Carbon on 10 November.
The framework says of AI: “It is a layered industrial stack — and at any given moment, the binding constraint sits at a different layer than it did six months ago.”
As we read it, Musk’s post places xAI’s constraint for 4.8 in the software layer, not in compute: by his account the rewrite took extra time and the SFT data needed cleanup. The framework predates Musk’s post and does not discuss xAI.
What Is Not Established
No Grok 4.8 benchmark figures have been published by xAI in the material we read, and Musk gave none. His 14 September post put 4.8 at 2.5 trillion parameters; the 10 November post does not restate a size. The post does not say how large the soft rollout is, whether Grok API customers get 4.8, or what it will cost.
We cannot verify the stack rewrite independently: it rests on Musk’s post. The 57-day figure measures the gap between his two posts; neither post gives the date when 4.8’s RL run began.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Musk says xAI has replaced JAX and SGLang with its own C/C++ code for RL and inference, after saying in September that 4.8 was trained on a new C++ stack. Grok 4.8 is in RL 57 days after he said it would start RL that week, by our arithmetic, with a soft rollout to Grok Bot users promised for this week and no published scores.
A note on sourcing. Musk’s statements are from his own posts on X of 14 September and 10 November 2026; Grok 4.7 details are from xAI’s launch page; the 2023 stack description is from xAI’s Grok announcement; JAX and SGLang descriptions are from their GitHub pages. SpaceX, Anthropic, Google and Similarweb figures are from our earlier pieces linked above. Day counts are our arithmetic. We have not seen any Grok 4.8 evaluation results.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
