No single investor led Instinct's new $1 billion round, which values the AI-assistant maker at $10 billion
· 3 min
Instinct raised $1 billion in a round with no named lead investor, a month after a round that valued it at a quarter of today's price.
L5 Tooling and infrastructure
The plumbing between a model and a product. Serving, retrieval, agent protocols, evaluation.
What this layer is
The layer nobody outside the industry can name and nobody inside it can avoid. Routing, caching, cost control, memory and the protocols agents use to talk to tools.
How it moves
Sits on models, carries applications. It repriced hard this year because on long agent runs the bottleneck stopped being model access and became everything around it.
The numbers
- $17.5B Fireworks AI valuation on $1B+ annualised revenue Jul 2026
- $13B Baseten valuation, 160% step-up in five months Jun 2026
- $1.15B Together AI Q2 bookings Q2 2026
- 2026-07-28 First breaking MCP spec revision since Nov 2025 modelcontextprotocol.io
Who is in this layer
- Fireworks AI: Inference serving platform. Repriced sharply upward as long agent runs made routing and caching the real bottleneck.
- Together AI: Inference and training cloud, raised at $8.3 billion on $1.15 billion of quarterly bookings.
- Baseten: Raised at $13 billion five months after a round at $5 billion, without disclosing revenue.
- Hugging Face: Where open models are shared. Nvidia agreed to buy it on 3 September 2026, which puts the company that makes the chips in charge of the place the open models live.
- Model Context Protocol: How agents talk to tools. Took its first breaking change since 2025 to handle agents that are long-lived rather than short-lived.
Stories in this layer
Two companies say they expect to spend a billion each putting auditors inside the lab
· 5 min
Anthropic and Accenture are funding evaluators with access "comparable to an employee's". Read the verb: expect to invest is not money spent, and not money committed.
What has happened here
No single investor led Instinct's new $1 billion round, which values the AI-assistant maker at $10 billion
· L5 Tooling and infrastructure · $1 billion · Confirmed
Instinct, a startup building a personal AI assistant, raised another $1 billion from investors Sequoia Capital, Benchmark Capital, and Coatue, according to the company's own announcement. Unlike many large funding rounds, none of the three is named as the lead investor; the release lists them as co-investors. The new round values Instinct at $10 billion, though that number appears only in the announcement's headline, not spelled out again in the body text. This is new money for a consumer AI product most people have not used yet, not a sign the industry overall is changing, but it is one of the larger single checks written for an AI startup this week.
A startup that answers phone calls for apartment managers and medical offices raises $350 million, now worth $4 billion
· L5 Tooling and infrastructure · $350 million · Confirmed
EliseAI, which sells AI software that answers phone calls and schedules tasks for apartment leasing offices and healthcare practices, raised $350 million at a $4 billion valuation. The round was led by Andreessen Horowitz and Bessemer Venture Partners, with Ontario Teachers' Pension Plan and others also putting in money. This is the kind of AI that is already answering real phone calls for property managers and medical offices today, not a research project, so if you've called a leasing office recently, there's a decent chance you spoke to software like this without knowing it.
Research paper site arXiv limits submissions to two a month as AI-assisted papers flood in
· L5 Tooling and infrastructure · 2/month cap; 40,363 Sept submissions · Confirmed
arXiv, the website where scientists post research papers before formal publication, is now limiting each person to two new submissions a month, after AI tools helped nearly double the number of papers submitted over two years: 40,363 in September 2026, up from 20,569 in September 2024. arXiv says AI is making it easier to produce "low-value papers" faster than its volunteer moderators can check them. It's a small, unglamorous sign of a much bigger pattern: AI is speeding up how fast people can produce written work of all kinds, and the places built to handle human-paced output are having to adapt.
Nvidia launches an open platform meant to stop AI agents that step outside their limits
· L5 Tooling and infrastructure · OpenShell software; Sentry on BlueField-4 chips · Confirmed
Nvidia announced its Open Agent Safety Platform on 28 September, open software and a reference hardware design for keeping AI agents inside the limits their operators set. One part, Sentry, runs on Nvidia's BlueField-4 data processors and watches agents from outside the software they run in; Nvidia says it can quarantine an agent that tries to move outside its boundary within milliseconds. Nvidia names Anthropic, Cisco, CrowdStrike, Microsoft, Hugging Face and SpaceXAI among those working with it, and says SpaceXAI uses it for Cursor coding agents and Grok models. For a company letting AI agents act on its systems, this is a tool meant to put a hard stop on an agent that goes past what it was allowed to do.
Anthropic and Accenture each say they expect to invest at least a billion dollars to put outside auditors inside AI labs
· L5 Tooling and infrastructure · each expects to invest at least $1bn over five years · Confirmed
Anthropic and the consulting firm Accenture announced a partnership on 18 September under which, in their own words, each expects to invest at least 1 billion dollars in building capacity in this area over the next five years. The money funds what Anthropic calls embedded evaluation: giving independent outside evaluators access comparable to an employee's inside its own labs, including watching models take shape in training and speaking directly to staff. Anthropic says it will fund Accenture's work directly, while a separate pilot with nonprofit evaluators runs on those evaluators' own money. Read the verb carefully. The companies say they expect to invest, which is not money spent and not money contractually committed, and Anthropic's release does not pace the spending across the five years.
Read the story: Two companies say they expect to spend a billion each putting auditors inside the lab · Anthropic
Baseten moves tool calls onto its own servers so AI agents stop making a round trip for every web search
· L5 Tooling and infrastructure · 15% lower end-to-end latency, company benchmark · Confirmed
Baseten, a company that hosts AI models for other businesses to run, has launched Hosted Tools, which lets a model call something like a web search from inside Baseten's own inference stack instead of handing the request back to the application and waiting for an answer to come back. Baseten says its own benchmarks show a 15 percent cut in total response time against the older method, a company figure with no independent test behind it yet. The first tool is web search, and rather than build one, Baseten has plugged in four existing search providers: Exa, Keenable, Parallel and You.com. This is the kind of plumbing that makes an AI assistant feel quicker when it has to go and look something up, without you ever seeing the change.
A startup founded by former Anthropic and METR staff raises $40 million to certify AI agents for insurers
· L5 Tooling and infrastructure · $40M Series A, $55M total · Confirmed
AIUC, short for Artificial Intelligence Underwriting Company, raised 40 million dollars in a Series A led by Ribbit Capital with participation from First Harmonic, which takes its total funding to 55 million dollars alongside an earlier 15 million dollar seed. One founder was Anthropic's first product hire; the other was a partner in McKinsey's insurance practice and chief operating officer of the AI safety research group METR. The company has built a standard called AIUC-1 that tests how agents hold up against risks including jailbreaks, hallucinations and data leaks, using 5,000 risk and attack combinations. Its own release names Cursor, ElevenLabs, Harvey, KPMG, Lovable, UiPath and Fin as companies that certify against it. For most people nothing changes today; this is an attempt to build the kind of inspection standard insurers and large companies want before they rely on AI agents more heavily.
Perplexity rebuilds its own database and cuts typical read delay more than fivefold
· L5 Tooling and infrastructure · median read latency 31.4ms to 5.60ms · Confirmed
Perplexity, the AI search company, said on 14 September that it built its own database, CobbleDB, to replace the off-the-shelf system it was using, and that median read latency fell from 31.4 milliseconds to 5.60 milliseconds, with the slowest one percent of reads falling from 123 milliseconds to 24.2. It says the new system is at least 20 percent cheaper at every pricing tier than what it replaced. It also says it was built by two human engineers and hundreds of AI agents in two months, a claim about its own process that no outside party has checked. This is Perplexity's own plumbing under its own product, not something anyone can use, but faster and cheaper backend infrastructure is the sort of change that shows up later as a quicker product or a lower price without either being announced.
An open source project raises serving speed for a major open AI model by up to 2.8 times
· L5 Tooling and infrastructure · 2.2x to 2.8x throughput versus vLLM's own prior release · Confirmed
Engineers working on vLLM, an open source system for running AI models, said on 13 September that they raised serving throughput for Moonshot AI's Kimi K3, an open-weight model, by 2.2 to 2.8 times against vLLM's own previous release, with latency down 56 to 60 percent over the same comparison. The work was done by engineers from Red Hat, Huawei, Inferact and Nvidia and published on vLLM's engineering blog. Several authors work at companies with a commercial stake in the hardware or software involved, so this desk treats the figure as a real but self-measured result rather than an audited one. For anyone running large open models on their own hardware instead of paying a lab by the token, a gain like this means more users served on the same machines.
Together AI cuts the price of training AI models on its platform by up to 70 percent
· L5 Tooling and infrastructure · 30 to 70% lower · Confirmed
Together AI, a company that rents out computing power for building and running AI models, lowered the price of fine-tuning, the process of further training an existing model on a company's own data, by 30 to 70 percent across most of the models it supports. For one model, OpenAI's open-weight gpt-oss-20b, the standard training price fell from 1.50 dollars to 0.40 dollars per million tokens processed. Together AI says the reduction comes from its own optimisation of the systems that do this training, not from a change in what the underlying chips cost. For a software team or startup that fine-tunes models to fit its own product, this is a direct reduction in one of the recurring costs of building an AI product, though it applies to training a model rather than running it once it is finished.
Together AI's open source code reaches more than 22 petaflops per GPU on Nvidia's Vera Rubin NVL72
· L5 Tooling and infrastructure · 22+ petaflops per GPU · Reported
Together AI ported its open source ThunderKittens programming tools to Nvidia's newest chip, called Vera Rubin NVL72, and says it reached over 22 petaflops of real-world performance per GPU, a measure of raw calculation speed, competitive with Nvidia's own closed-source software for the same job. A petaflop is a thousand trillion calculations per second; Nvidia's previous generation chip, the B200, tops out at 9 petaflops per GPU on this kind of workload on paper, so Vera Rubin represents a large jump in raw hardware capability. This is Together AI's own benchmark of its own open source code, not independently verified, and it is a technical achievement inside the AI infrastructure world rather than a product change that affects prices today. For anyone using an AI product, tools like this are part of what eventually lets AI companies use new chips at close to their full speed once they are available.
An AI slide-deck generator cuts its own costs by about 90 percent with a specialised model
· L5 Tooling and infrastructure · about 90% cost cut · Reported
Fireworks AI and Genspark, a company that generates presentation slides with AI, released a specialised model called Gen-1 Slides that they say produces decks of similar quality to Anthropic's Claude Opus 5 at about one seventeenth of Opus 5's input-token price, and cut the share of decks rated low quality in production from 18 percent to 3.6 percent. For a team producing 1,000 decks a month, the companies say that amounts to roughly a 90 percent cut in cost, from about 4,200 dollars to 400 dollars. These are the companies' own comparisons on their own workload, not independently verified. If costs like this keep falling across the industry, it is one reason AI-generated documents and presentations are likely to keep getting cheaper and more common at work.
Together AI launches a cheaper GPU rental tier at half the price of its regular rate
· L5 Tooling and infrastructure · 50% of on-demand rate · Confirmed
Together AI, a cloud company that rents out GPU computing power, launched a new pricing tier called preemptible compute, priced at a flat 50 percent of its regular on-demand rate. The tradeoff is that Together can reclaim the computer at short notice, giving five minutes' warning before it does. This is a new, cheaper option alongside Together's existing prices, not a cut to what existing customers already pay. For a developer renting GPU computing power to train or run an AI model, this is a real way to cut costs, if the work can tolerate being interrupted.
Baseten cuts the time to swap in an updated AI model on a live server to under 40 seconds
· L5 Tooling and infrastructure · under 40 seconds · Reported
Baseten, a company that runs AI models on servers for other businesses, said it can now update a large model like GLM-5.3 across independent computer clusters around the world in under 40 seconds, with only 6 seconds where incoming requests have to wait. It does this by sending only the parts of the model that actually changed, compressing what would otherwise be 716.56 gigabytes of updated data down to 1.55 gigabytes. These are Baseten's own reported figures for its own system, not independently benchmarked, but they describe a real engineering change rather than a new product. For anyone using an AI powered app or chatbot, this is part of why those services can update to a smarter model without going offline while it happens.
An open source server runs a million token session on one eight GPU node
· L5 Tooling and infrastructure · 1M context · Confirmed
A million tokens is roughly a 700,000 word conversation, about seven long novels. Software that runs AI models can now hold one that long on a single machine, so the cost of AI that works for hours on one task just fell without anyone buying new hardware.
Nvidia agrees to buy Hugging Face, the home of open models
· L5 Tooling and infrastructure · $12.93B · Confirmed
Nvidia makes the chips nearly all AI runs on. Hugging Face is the site where researchers publish and share their AI models freely. The company that makes the engines has bought the garage everyone parks in.
Hugging Face publishes an initial 207 kernels for running models in the browser
· L5 Tooling and infrastructure · 207 kernels · Confirmed
These are small programs that let an AI model run inside a web browser, using the graphics chip in your own computer rather than a server elsewhere. They work only where the browser, the operating system, the graphics chip and its driver all support it.
MCP ships its first breaking spec revision since November 2025
· L5 Tooling and infrastructure · Stateless · Confirmed
MCP is the standard that lets AI agents use outside tools. Its first breaking change since November 2025 removes the setup handshake, so every request now carries its own version and permissions. Anything built on the old session behaviour has to be rewritten.
Baseten raises $1.5B across tranches priced at $11B and $13B, five months after a round at $5B
· L5 Tooling and infrastructure · $11B and $13B · Confirmed
Baseten runs AI models for other companies. It raised 1.5 billion dollars in two parts, priced at 11 and 13 billion, five months after a round that valued it at 5 billion. No revenue figure was disclosed with either round.
Strata, the whole AI stack, explained simply. Every number carries a source and a confidence label.
Today · The Stack · Stories · Archive · Glossary · Break the Chain · Focus · About · RSS feed
Privacy · Terms · Corrections · Report an error
Also from Dheeco: Dheeco · AP Fact Check
Published by dheeco.com · © 2026 Dheeco