AI Startups & Viral AI (Oct 12): Decart, Vinci & Qwen
AI startups news for Oct 12: Amazon nears a $7B deal for Decart, Vinci raises $250M for chip simulation, and Microsoft ships a Qwen model 58% under OpenAI.
Amazon is in advanced talks to buy Decart for around $7 billion, the Wall Street Journal reported on Sunday, a month after Anthropic did due diligence on the same company and walked away. Decart does not sell a frontier model. It sells software that makes models cheaper to run and lets them move between chips built by different manufacturers, and AWS has been a customer since 2025.
That is the shape of this week's money. AI startups announced about $2.39 billion across 14 rounds of $20 million or more between October 6 and October 12, according to AI Funding Tracker. The loudest cheques went to the layer underneath the model: portability, simulation, and security testing. The viral argument of the day was about what cheap models do to the companies selling expensive ones, and what ten minutes with an assistant does to a person.
AI startups: three bets on what the model runs on
Amazon and Decart: $7 billion for the software that makes silicon fungible
Decart is an Israeli AI company founded in 2023 by Dean Leitersdorf, Orian Leitersdorf and Moshe Shalev. It raised $300 million in May at roughly a $4 billion valuation, in a round led by Radical Ventures with Nvidia participating, taking total funding to about $450 million. In August, Anthropic explored a deal near $6 billion and became the preferred bidder; in September it withdrew after diligence. Amazon is now at the top of that range. Source: Reuters
The product is the reason. Decart's core software optimises how models execute and lets a workload move between accelerators without a rewrite. AWS gave the company early access to Trainium3 in 2025 and is already a strategic customer. For a cloud that designs its own silicon, buying the portability layer is cheaper than persuading every customer to re-platform by hand.
What this changes for someone shipping today: if portability becomes a first-party feature of a big cloud, the premium you pay for being locked to one accelerator vendor starts to look like a choice rather than a constraint. The risk is that talks can still collapse, and neither company has confirmed a price.
Vinci: $250 million to simulate the chip before anyone builds it
Vinci is a Palo Alto company that sells physics simulation for chip and hardware design. Reuters reported on October 6 that it raised $250 million at a $1.5 billion valuation, co-led by Advent, Temasek and Xora, with AMD Ventures, Eclipse, Khosla Ventures and Madrona joining. Source: Reuters
The company is about 70 people and previously raised $36 million. It started with thermal simulation and plans to add vibration and electromagnetics, spending the new money on compute and simulation engineers. CEO Hardik Kabaria told Reuters the current focus is growth rather than margin.
Every custom accelerator programme is gated by how long it takes to learn whether a design will thermally survive, and that loop currently runs through fabricated silicon. Simulation that shortens it is a tax cut on the cost of competing with Nvidia, which is why AMD's venture arm is on the cap table.
Hadrian: $40 million to attack the agents you have already shipped
Hadrian is an Amsterdam company whose platform tests systems the way an attacker would, aimed at AI agents that hold credentials and can act on production infrastructure. Its $40 million Series B was co-led by Forgepoint Capital International and SmartFin, with HV Capital, Motive Partners and Picus Capital participating, bringing total funding to $65 million. Its previous round was a 10.5 million euro seed led by HV Capital in 2022. Source: TechStartups
The judgement here is about timing rather than size. Agent deployments moved from demos to payroll and finance workflows in 2026, and the testing tools for them are still mostly prompt-injection word lists. A $40 million round will not build the category, but it is the first serious European cheque aimed at testing agents adversarially rather than testing models.
What this changes for someone shipping today: if an agent has write access to a system of record, your security review has to include an attacker who talks to it in natural language.
Startup Snapshot
| Startup | What They Do | Latest Round | Valuation | Why It Disrupts |
|---|---|---|---|---|
| Decart | Software that cuts model running costs and moves workloads across chip vendors | ~$7B acquisition talks with Amazon (WSJ) | ~$4B in May 2026 | Makes accelerator switching a cloud feature, weakening vendor lock-in |
| Vinci | Physics simulation for chip and hardware design | $250M Series B | $1.5B | Shortens the design loop that gates every non-Nvidia accelerator |
| Hadrian | Adversarial testing platform for agentic AI systems | $40M Series B | Not disclosed | Turns agent security from prompt lists into an attack surface |
Viral AI: a cheaper token, a wider gap and a lost habit
Microsoft is reselling Alibaba's model at $0.042 a million tokens
Microsoft's Decision-1 shipped on October 9 at $0.042 per million input tokens with output free, against the $0.10 OpenAI had published three days earlier, a 58% cut below its own partner. It runs on Alibaba's Qwen3.5-9B, handles 32,768 tokens of context, is text only, and returns calibrated probabilities as JSON rather than prose. Source: MarkTechPost
Microsoft also published a head-to-head accuracy table, reporting 83.5% for Decision-1 against 79.4% for OpenAI's GPT-6 Luna Decisions across 36 benchmarks it ran itself. A vendor grading its own model on its own harness is a claim, not a measurement.
The part that travelled fastest is the sourcing. A Microsoft product built on Chinese open weights, priced to undercut the model partner Microsoft is closest to, tells you the open-weights ecosystem is now good enough to be a supply chain rather than a hobby. What could make this read wrong: a decision-scoring model is a narrow task, and narrow tasks are exactly where a small fine-tune can win.
Karpathy says we no longer agree on what these models can do
Andrej Karpathy, a founding member of OpenAI and former director of AI at Tesla, posted that the gap in shared understanding of LLM capability is widening. The Hacker News thread it anchored collected 379 points and 537 comments in under a day, a comment-to-point ratio of 1.4 that usually signals an argument rather than agreement. Source: Hacker News
Read it beside the Microsoft table above and the day gets legible. Leaderboards disagree because they weigh different things, vendors publish their own harnesses, and effort settings inside one model swing cost by 70%. Practitioners who use these systems daily and people who read about them are no longer looking at the same object.
What this changes for someone shipping today: stop quoting a single capability number in a build decision. Pick the benchmark that matches your workload, run it yourself, and record the effort setting and cost per task alongside the score. Source: AIToolsRecap
Berkeley: ten minutes with a chatbot, and less persistence afterwards
Researchers at UC Berkeley ran three randomised trials with 1,222 participants. The group given ChatGPT performed better while they had access to it, then solved fewer problems and skipped more once it was removed; the control group held steady or improved. The work was presented at COLM in October. Source: Berkeley News
Ten minutes is the detail worth sitting with. This is not a study about a generation losing the ability to write; it is a measurable change in behaviour after a very short exposure, and it implies the effect is a property of the interaction rather than of the person.
What this changes for someone shipping today: remove friction from a task and you also remove the reps that made the user able to do it unassisted. Teams selling to professionals should expect a harder renewal conversation in two years, not in five.
Viral Snapshot
| Story | What Went Viral | Engagement Signal | What's Actually True |
|---|---|---|---|
| Microsoft Decision-1 | A Microsoft model on Alibaba's Qwen at $0.042 per million tokens | 58% under OpenAI's published price | Pricing and benchmark figures are vendor-published, not independently run |
| Karpathy on capability gaps | A post arguing shared understanding of LLM capability is splitting | 379 points, 537 comments on HN | The framing is opinion; the disagreement in the thread is observable |
| Berkeley persistence study | Ten minutes of AI use changed later behaviour | Three trials, 1,222 participants, COLM | Randomised trials, but 1,222 participants is a modest sample for a habit claim |
What this means for builders
Put the two sections together and there is one argument, and it is a margin argument.
Every startup story today and this week is capex-adjacent. Amazon wants Decart so AWS can move workloads onto its own silicon. Vinci wants to shorten the interval between a design and knowing whether it works. Hadrian sells insurance on the fact that agents with credentials are now normal. None of them sell intelligence. They sell the cost of using it, and cost is the only variable in this market that has fallen every single day since the year's start.
The viral section is the same story from the other end. Microsoft undercut its closest partner by 58% using Chinese open weights, which is what happens to a supplier when the input it resells becomes a commodity. Karpathy's thread is the noise of a market where nobody can agree on the price of capability because the price keeps changing and the ruler keeps bending. Berkeley is the reminder that the buyer is a person whose ability to work without the tool is itself degrading, which caps how much of the value you can capture.
The practical read for a founder building on these APIs: your defensibility is not the model call. It is the workflow, the data you accumulate, and the switching cost you build that is not vendor lock-in. The startups raising today are being paid to reduce the cost of the thing you resell. If your plan assumes inference stays expensive, this is the week the plan broke.
Related reading
- Inference chips that beat Nvidia on price, where the silicon argument stood in August.
- A chipmaker taking on Nvidia, and a 1B inference chip bet, the earlier round of challenger funding.
- Korea's chip law and frontier compute policy, the government side of the supply chain.
- Frontier labs pacing training against cyber risk, the security thread Hadrian is monetising.
- AI that raised homework scores while safety gaps widened, on what AI use does to the user.
Frequently Asked Questions
What are the top AI startups to watch today?
Amazon's talks to buy Decart for around $7 billion, Vinci's $250 million round for chip simulation, and Hadrian's $40 million for agent security testing. All three sell the cost of running models rather than the models themselves, which is where this week's $2.39 billion in disclosed AI funding concentrated.
Why is Amazon buying Decart instead of building the software itself?
Because Decart's portability layer lets AWS move workloads onto its own silicon without asking customers to rewrite them. AWS is already a customer and gave Decart early access to Trainium3. Buying that switching tool is faster than rebuilding it, and it directly reduces Nvidia's share of the compute AWS rents out.
Is the Karpathy claim about LLM capability gaps just an opinion?
Yes, it is an opinion, and the evidence for it is the argument it produced. The thread drew 537 comments against 379 points. The measurable version of the claim is that leaderboards disagree, vendors grade their own models, and effort settings inside a single model change cost by up to 70%.
Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch ->
