If you pilot NeuReality's NR-NEXUS software or await its NR2 silicon, your vendor will become the engineering arm of a company with no shipping product, and nobody has said what happens to your contract. Unconventional AI, the Mountain View startup run by Naveen Rao, signed a definitive agreement on October 26 to acquire NeuReality in a cash-and-stock deal with undisclosed terms. Both boards have approved it and it is expected to close in Q4 2026. The announcement talks about people and a 1,000x efficiency goal. It says nothing about customers.
That silence is the whole reason to act now. Between signing and close, you still have a counterparty with an incentive to keep revenue and goodwill intact. After close, you are a line item in an integration plan written by a company whose stated aim is a different kind of computer.
What Unconventional AI Is Actually Buying
Unconventional AI is buying a team that can take silicon to production, and every public statement about the deal frames it that way. Ynet reports that NeuReality's engineering staff, including specialists in system architecture, semiconductor design and AI software, will join Unconventional, along with CEO Moshe Tanach and the founding team, and that engineering stays in Israel. NeuReality is headquartered in Caesarea.
Rao told Reuters that about 40 NeuReality employees will join and that his company now needs to "build up the production bones" after laying its research foundation. "Things have gone a lot faster than anticipated," he said. The same report says Unconventional had planned on three years of R&D and five years to a product, and has already built a first test chip and nearly finished a second. Reuters also notes that Rao sat on NeuReality's board, so the buyer had an inside view of the company it is acquiring.
The price is not public. Calcalist reports the deal is expected to return more than the roughly $105 million NeuReality raised from Arm, Samsung Ventures, OurCrowd and others. Bizportal, citing Globes, puts it at an estimated $150 million to $200 million, against a PitchBook valuation of about $220 million in 2023. Treat both numbers as press estimates.
The buyer has money. Unconventional raised a $475 million seed round at a $4.5 billion valuation in December 2025, led by Andreessen Horowitz and Lightspeed, as the first part of a raise that could reach $1 billion. Rao has described the goal as a computer that is "as efficient as biology." Ynet is explicit that the 1,000-fold improvement is an objective, not a demonstrated result, and that no timetable for commercial production has been given.
What NeuReality Sells Today
NeuReality sells three things, and none of them is mentioned in the deal coverage. Its own site, checked October 30, still lists them with no acquisition notice:
- NR-NEXUS, an "inference operating system" that manages serving, routing, utilization, latency and cost across GPU and XPU clusters. It is sold three ways: a serverless API billed per token, a managed deployment with enterprise SLAs, and an annual licence for managed or owned infrastructure.
- The NR2 AI-SuperNIC, networking silicon for scale-out GPU clusters at 1.6 Tbps with Ultra Ethernet Consortium support.
- The AI-CPU, an Arm-based chip that pairs a CPU with media processors and NeuReality's AI-Hypervisor.
An inference operating system, in NeuReality's usage, is the software layer that decides which accelerator serves which request and keeps utilization high across mixed hardware. It sits in the same part of the stack as the open-source serving layers compared in our llm-d, KServe and Ray Serve guide.
Two details make the timing awkward. NR-NEXUS launched in beta in March 2026, with no named customers and no published pricing at launch. Tanach pitched it as giving operators "flexibility rather than lock-in." And NR2 was promised to select customers in the second half of 2026, with mass production ramping in 2027, with the networking die shipping first as a standalone part called NR2n. The second half of 2026 ends in two months. Futurum's analyst listed customer adoption timelines as the first thing to watch when NR2 was announced in September 2025.
So the customers with the most exposure are the ones with the least leverage: beta users of a software product with no list price, and early-access buyers of a chip that has not reached volume.
Why Rao's Own History Is the Best Guide
The most useful precedent here involves Rao himself. Intel bought his first company, Nervana, for $408 million in 2016. After Intel bought Habana Labs for $2 billion, it ended development of the Nervana NNP-T training chip, a decision Network World reported in February 2020. For the NNP-I inference chip, Intel said only that it would "deliver on current customer commitments."
That phrase is about as much as an acquirer usually offers: existing orders get filled, and the next generation is open to cancellation.
The strongest case for the other side deserves a hearing. Unconventional is buying engineers who know how to tape out chips, which is the production experience Rao says his company now needs. A buyer that wants talent has every reason to keep those engineers happy, and shutting down the product they built is a fast way to lose them. NeuReality's software was also designed to be hardware-agnostic, which could make it useful as a control plane for whatever Unconventional eventually ships.
Compare that with how Groq handled its licensing deal with Nvidia last December. Groq's announcement said in plain words that GroqCloud "will continue to operate without interruption" and that Groq would remain an independent company under a new CEO, even as its founder moved to Nvidia. We covered why that structure matters in the Nvidia and Reflection AI piece. NeuReality customers have nothing comparable. Until they do, plan as if NR-NEXUS and NR2 are assets Unconventional may keep, fold in or retire.
The same question has come up in several infrastructure deals this year. When d-Matrix bought Wallaroo, the question was whether "any hardware" survived the deal. When AMD bought Taalas, it was which model a customer would freeze into silicon. And when Cloudflare bought Deno, Deno Deploy got a six-month shutdown notice.
What Is at Risk in Each Contract Type
Your exposure depends on which of NeuReality's three NR-NEXUS paths you are on, because each fails differently if the product is wound down.
On the serverless API you pay per token and own nothing, so the risk is an endpoint that stops answering. Switching cost is low if your code calls an OpenAI-compatible interface, and the work is keeping a second provider warm. Our Together, Fireworks and Groq comparison covers alternatives for open-weight models, and Groq and Baseten both run hosted inference.
On a managed deployment, NeuReality runs dedicated capacity under an enterprise SLA. That SLA is only as good as the team behind it, and that team is moving to a company focused on new silicon. Ask who will be on call after close.
On an annual licence you run NR-NEXUS on your own or managed infrastructure. This is the most exposed position, because the software sits in your serving path and depends on NeuReality for patches. A licence with no source escrow and no end-of-life notice period leaves you running an orphaned control plane.
For NR2 early-access buyers the question is simpler: will the part you designed a cluster around reach mass production in 2027, and on what terms? The NR1 got a software upgrade for UEC 1.0 support. Whether NR1 and NR2 will keep getting firmware after close is the kind of thing that only appears in writing if you ask for it.
What to Do Before the Q4 Close
You have a few weeks while your terms are unchanged and the seller still wants the deal to close without customer complaints.
This Week:
- Pull every NeuReality order form, licence and pilot agreement and find the change-of-control and assignment clauses. Our review of AI vendor exit clauses found data-export windows ranging from 0 to 90 days, so check what yours says.
- Email your NeuReality account contact and ask four questions in writing: Will NR-NEXUS remain available after close? Will the serverless and managed services keep running, and for how long? Will NR2 ship to early-access customers on the published schedule? Who supports NR1 firmware?
- List every workload that routes through NR-NEXUS today and note which ones would break if the endpoint or licence server went dark.
This Month:
- Ask for a signed continuity letter covering at least 12 months of support, security patches and service availability, plus a minimum notice period before any end-of-life.
- For licensed deployments, request source-code escrow or a perpetual licence to the version you run. A team being absorbed into another company is the moment a vendor is most willing to grant either.
- Stand up a fallback serving path on an open-source stack or a second hosted provider, and run a fraction of real traffic through it so the switch is tested before you need it.
Before Renewal:
- Do not renew or expand an NR-NEXUS commitment until the buyer's product plan is public. If you must sign, keep the term short and make termination for convenience mutual.
- If you are an NR2 design partner, tie any purchase commitment to delivery milestones and a refund if mass production slips past 2027.
The Bottom Line
Unconventional AI is paying for engineers who can turn a research chip into a production one, and on its own terms that is a sensible purchase for a $4.5 billion company with a test chip and a 1,000x goal. Whether the products those engineers built survive is a separate decision, and the buyer has not made it public. The last time Rao was on the other side of a chip deal, Intel ended the Nervana training chip and promised only to meet existing customer commitments on the inference one. NeuReality customers should get the same promise in writing before close, plus a notice period and an exit. Send the four questions to your account contact this week.
Continue Reading
- d-Matrix Bought Wallaroo. Get 'Any Hardware' in Writing.
- AMD Bought Taalas. Now Name the Model You'd Freeze.
- Nvidia Could Buy, Invest in or Acquihire Reflection AI
- Five AI Vendors Give Customers 0 to 90 Days to Get Data Out
- Deno Deploy to Shut Down Six Months After Cloudflare Deal
- KServe and llm-d Beat Ray Serve for Kubernetes Teams