Olix triples its valuation to $3.3B, betting the inference bottleneck is silicon, not models

Good morning 👋 A chip startup just tripled its own price tag in six months, and the first customer won't see silicon for another year and a half. Below: why investors are buying the roadmap anyway.

In today's issue:

• 🔭 Olix triples its valuation to $3.3B, betting the inference bottleneck is silicon, not models
• 🧠 Sakana turns an open Kimi model into a Japan-only business API
• 🔬 Cheap fixes for two habits agent stacks didn't know were expensive
• 🛠️ Put a real number on local voice AI vs. your ElevenLabs bill
• 🔗 OpenAI's CFO wants a new way to price a model's worth

🔭 THE ONE THING

⚡ Olix triples its valuation to $3.3B, betting the inference bottleneck is silicon, not models

Olix raised $312M at a $3.3B valuation, up from just over $1B in February, with Fundomo leading and Arm, Hudson River Trading, and Reed Hastings writing checks alongside existing backers doubling down. The pitch is DX-1, a decode accelerator that claims over 10,000 tokens per second per user on 100B-parameter models using on-chip SRAM instead of the HBM everyone else is scrapping over. That's a real architectural bet, not a spec-sheet flex, which is probably why Arm and a high-frequency trading shop both wanted in. First customers don't see silicon until H2 2027, so this valuation is built entirely on a roadmap: six months, 3x the price tag, zero shipped chips. I want benchmarks against actual Nvidia hardware before I call this the inference chip.

🧠 MODELS & RELEASES

• 🐟 Sakana AI launched Namazu, an API that takes Moonshot's open Kimi K2.6 and tunes it for Japanese business use, with web search and code execution baked in. It's OpenAI-compatible, so switching over is a base_url swap and a new key; Sakana says the tuning pushed its FairPoliticsQA benchmark from 34% to 56%. Sakana AI

🔬 RESEARCH HIGHLIGHTS

• TokTier found a dumb tax baked into agent serving: front ends re-tokenize the entire transcript on every call, even when the agent only appended a small tool result. Across 154K real agent calls, that re-tokenization step ate up to 64% of time-to-first-token. TokTier tokenizes just the new window and splices it onto the cached state instead, dropping median TTFT 16-34% under vLLM. If your agent loop resubmits a growing transcript every turn, this is the kind of fix that just shows up as a smaller bill. paper

• Zero-Mem cuts out a habit most agent memory systems have picked up: spending a separate LLM call to write and retrieve memories, which also tends to paraphrase away the original evidence. It keeps full interaction history in an entity graph plus a timeline, then filters deterministically at query time with no model calls in the retrieval path. On long-memory benchmarks it matched the strongest existing method's accuracy while cutting 57.6% of its time. paper

• Sycophancy costs more than tone points. Researchers ran vision-language models through a "spot the difference" task where each model sees a different image and has to argue it out over dialogue. Models routinely folded, agreeing the images matched even when their own image plainly disagreed. A steering vector trained to suppress sycophancy cut those errors. Worth remembering the next time a multi-agent pipeline has one model "confirming" another's read. paper

🛠️ TRY THIS

Put a number on the local-vs-API question, using this week's tool

1. Clone debpalash/OmniVoice-Studio and point it at whatever GPU you already have, even a rented spot instance if you don't own one.

2. Run the same job you'd normally send to ElevenLabs: clone a voice, dub a short clip, or dictate a paragraph.

3. Time it. Log $/minute against your last ElevenLabs invoice for that same task.

4. If local wins on cost and the quality holds, move the recurring work off the API. If it doesn't, at least you're arguing from a number instead of a hunch, which matters while everyone from Olix on down is repricing what inference actually costs.

Prompt: Write a Python script that times OmniVoice-Studio's voice-clone inference on a sample WAV file, tracks GPU memory used, and outputs a $/minute figure given my hourly compute cost.

Worth a look

• OmniVoice-Studio (github.com/debpalash/OmniVoice-Studio) bundles voice cloning, dubbing, dictation, and an audiobook maker into one local stack. 9,455 stars in, and it's the closest open answer to ElevenLabs if you'd rather not pay per character.
• Raven (github.com/EverMind-AI/Raven) is a memory-first agent harness built on EverOS, with MiroThinker handling the deep-research and reasoning side. Early (3,486 stars), and worth a look if you're building agents that need to remember something past the current session.

🔗 QUICK LINKS

• OpenAI CFO Sarah Friar says stop grading models on price per token and start grading them on total cost to finish the job, retries and errors included. Building abundant intelligence

That's the lineup. Back tomorrow.

Pradeep Perugu