ANVE: the private AI office, on the record. See what that means

ANVE

Engineering note 01 · 24 August 2026, published 4 September 2026

Pricing a private AI appliance from the BOM up.
Three of five tiers were underwater.

ANVE is a private AI appliance for teams whose work can't leave the building. Before we tooled anything we priced it the way most hardware startups don't: from the bill of materials up, at Indian street prices, on the day. This note is what that exercise found, including the mistake we had to retract. Prices below were checked on 24 August 2026 and will have moved.

First ladder5 tiers, 3 below cost
Main causeDRAM up ~4× since 2025
Second causeGeForce EULA §2.8
01

The first ladder was priced from aspiration

We started where founders usually start: what would a small firm, a mid-size firm and an institution pay for a box that runs meetings, documents, code and a private GPT locally? That gave a five-tier ladder from about ₹95,000 to ₹15 lakh plus custom. Then we costed each tier at landed Indian prices (GPU, CPU, board, memory, storage, chassis, import duty, freight) and compared. Result: the ₹95,000 tier landed at roughly ₹1.54 lakh, the ₹6 lakh tier at roughly ₹10.5 lakh, and the ₹12 lakh tier at roughly ₹22.8 lakh. Only one tier had margin. A ladder priced top-down is a wish list.

Update, 7 October 2026. We have retired the ₹95,000 entry target: our platform has never been tested on a Ryzen AI 300 part, and at that price the hardware margin was about 13%. The lineup is now ANVE One, Office and Office Pro, with indicative launch prices from ₹2,99,000 excluding GST; final prices are confirmed after our own engineering units are validated. See the lineup.

02

Memory did most of the damage

The 2026 DRAM shortage is the whole story of the small tiers. A 32 GB DDR5 kit that cost about $95 before the crunch was about $392 when we checked; 64 GB was about $849; a 128 GB two-stick kit was about $3,399. Density carries a premium of roughly 2×, so 128 GB as four 32 GB sticks (about ₹1.63 lakh) is far cheaper than two 64 GB sticks (about ₹3.25 lakh). Volume does not help here: in a shortage, volume buys allocation, not price. TrendForce was guiding a further 13–18% rise in Q3 2026. If you are speccing anything with a lot of RAM this year, that one line decides your margin.

03

The GeForce EULA took consumer GPUs off the table

We had assumed an RTX 5090 in the large tiers. Two problems. Street price in India was between ₹5.7 lakh and ₹7.4 lakh depending on stock, so two of them alone exceeded the tier's selling price. And the NVIDIA GeForce driver licence (nvidia.com/en-us/drivers/geforce-license, §2.8, rev. 25 Feb 2025) states the software "is not licensed for datacenter deployment". ANVE is pitched as a multi-user private AI server for enterprises and government, which is the deployment that clause is about. We read it at the primary source, not in press coverage, and we are not going to ship a product whose licence we would have to argue about. The RTX PRO line has no such clause.

04

The retraction

Our first revision claimed a complete 128 GB AMD Strix Halo system cost less than 128 GB of DIMMs alone, which would have been a remarkable inversion. It was wrong twice over: the system price we used was stale, and the DIMM comparison used the expensive two-stick configuration while the same document recommended the four-stick one. Against the configuration we actually specify, the complete system cost about ₹1.25 lakh more. What survives is a narrower and more useful claim: measured per addressable gigabyte, the 5090 was about ₹17,800 per GB of VRAM and Strix Halo about ₹2,240 per GB of unified memory, a 7.9× difference on the constraint that decides which models a box can run at all, traded against roughly 256 GB/s versus 1.8 TB/s of bandwidth, weaker concurrency, and a less mature ROCm stack. We will benchmark before we quote either number to a customer.

05

What we changed

The ladder was rebuilt bottom-up. The two small tiers moved to AMD unified-memory parts; the two large tiers moved to RTX PRO cards with a licence that permits the use; the ₹95,000 tier was first killed and then restored on a Ryzen AI 300 part with a far-field microphone array, because a 4B-class model on a 50 TOPS NPU can transcribe and summarise a meeting locally, and that is demonstrable in sixty seconds in a way a GPU spec sheet is not. Every tier now sells above landed cost. The prices you will eventually see are not the ones in the first ladder, and this note is why.

06

What is not real yet

There is no physical prototype. Every image of the appliance on this site is a labelled render of the enclosure we are building toward. We are taking refundable reservations (₹999 in India, US$12 elsewhere) to find out who wants this before we tool anything. A reservation guarantees no date, specification or certification, and it is refundable any time before you confirm an order in writing. The terms are two screens long and say so. Reserve one if you want to be on the list when the spec exists, or write to hello@anveai.com if you think the BOM reading above is wrong. We would rather find out now.