A 100T Transformer Model Coming? Plus ByteDance Saga and the Mixtral Price Drop

Rumors of a “GPT 4.5” release were met with unusually direct denials from multiple OpenAI employees, with one pointing to the pattern of a consistent hallucination and another warning that any “4.5” would not plausibly appear silently—especially not with a telltale API string. The exchange didn’t just cool hype; it also shifted attention back to concrete, market-moving developments: a claimed transformer-focused chip breakthrough, a sharp price collapse for Mixtral, and ByteDance’s alleged use of OpenAI technology to accelerate a competing model.

On the hardware side, a new company, H. (as referenced in the transcript), claims it has built the world’s first “Transformer supercomputer” designed from the ground up to run transformer workloads. The pitch is that the company “burned” the Transformer architecture into a custom chip—code-named “Soo”—so that each transistor can be optimized for the computations that dominate large language model inference, especially matrix multiplication. The company’s marketing centers on throughput gains—massively outperforming Nvidia H100 on tokens per second—and on enabling real-time interaction, such as voice agents that can ingest thousands of words in milliseconds and generate outputs without the typical pauses users experience.

That strategy echoes a broader bet described via an earlier article: two Harvard dropouts raised millions to design an AI accelerator dedicated to large language model inference, arguing that generalized hardware may not deliver the biggest gains. They acknowledge a key risk—specializing too narrowly as workloads evolve—but contend that if the bet pays off, the payoff could be dramatic: claims of up to 140× throughput per dollar and scalability toward extremely large models (up to 100 trillion parameters). The chip is reportedly slated for availability in 2024, with a funding round planned for early next year, though investors remain skeptical given the challenge of entering semiconductors.

Meanwhile, “Mixtral” is delivering an immediate, user-visible story: steep, rapid price cuts. The transcript traces a sequence of reductions after the model’s announcement—first to $2 per 1 million tokens, then down by 70% to 30 cents per 1 million tokens, and later to 27 cents per 1 million tokens—attributed to providers competing aggressively and even offering Mixtral for free in some cases. The implication is that performance improvements are arriving alongside faster cost declines, raising the question of where “intelligence per dollar” could land by the end of 2024.

Sebastian Bubeck—author of Sparks of AGI and associated with the F series—ties the market shift to a research goal: enabling reasoning at minimal scale. He frames the mission as identifying the smallest ingredients needed for GPT-4-like intelligence, not merely fitting large models onto phones, even though on-device models are trending upward.

Finally, the ByteDance saga adds legal and competitive pressure. The transcript alleges ByteDance secretly relied on OpenAI’s API (including for “Project Seed”) to develop foundational LLM code, despite OpenAI terms forbidding using model outputs to develop competing AI products. OpenAI reportedly banned ByteDance from ChatGPT, and ByteDance leadership is quoted as suggesting a “super strong” model may arrive soon, with open weights/open sourcing expectations left open as the timeline tightens. The overall picture: hype about GPT 4.5 cools, while hardware specialization, inference pricing, and competitive tactics accelerate the real race toward more capable—and cheaper—systems.