A mega funding round tells you almost nothing about whether to buy

A category you were quietly considering raises an enormous round, and for about a week the coverage makes the decision feel obvious. Robotics foundation models and AI-generated 3D have both been through this recently. The temptation is to read the number as a verdict: someone with far more information than you just concluded this works, so it must be ready. That inference is wrong in a specific and useful way. A venture round is a bet on where a company will be in seven years. Your decision is about whether a tool holds up in your workflow next Tuesday, and those two questions have almost no evidence in common.

This guide is for a solo operator, a freelancer or a small team deciding whether to put a paid AI tool inside work you deliver to someone else. It is not investment advice and it will not help you pick which robotics startup to back. If you have a procurement team and a lawyer, they will want more than this. What follows is the short version of what they would do: ignore the press release, read the pricing page and the terms of use, and price the work per finished unit rather than per month.

The wave is real and it is not a signal about your purchase

The money is not imaginary. Physical AI startups took $47.4 billion across 521 deals in the first half of 2026, against $12 billion across 470 deals in the second half of 2025 and $26.4 billion across 436 deals in the first half of that year [1]. For scale, the whole of 2022 through 2024 came to $41.9 billion combined [1]. That is a genuine repricing of the category.

It is also lumpier than it looks. One deal accounted for nearly a third of all venture dollars in the half: Waymo’s $16 billion Series D, raised in February at a $126 billion valuation [1]. A headline about “the robotics funding wave” is partly a headline about one self-driving company, and that has no bearing on whether a 3D asset generator will survive contact with your asset pipeline.

Sometimes the round does come with operating evidence attached. Meshy raised nearly $400 million at a $1.5 billion valuation in July 2026, and disclosed alongside it more than 12 million registered users, more than 100 million 3D models generated, annual recurring revenue growing about 12x year over year, and named customers including Nexon, NetEase Games, Bambu Lab and Hugo Boss [2]. Those numbers are worth something to you. The $400 million is not. Read past the check size to the operating figures, and if there are no operating figures, note that as the finding.

Whether you can buy it today is the first filter

Run this before anything else, because it disqualifies a surprising share of the companies in a funding wave. Go to the vendor’s site and try to find a price.

Genesis AI, one of the robotics foundation-model companies inside this wave, currently lists three things: Eno, described as its first general purpose robot; GENE-26.5, a foundation model the company says lets robots see, understand instructions, apply reason and manipulate with precision; and Genesis World 1.0, a simulation platform it says runs thousands of simulated trials in minutes [3]. There is no public pricing page, no API documentation and no self-serve sign-up. What there is is a waitlist, where you pick “Commercial” or “Home”, and a stated plan to begin targeted customer deployments by the end of 2026 [3].

Nothing about that is a criticism. It is the normal shape of a company building hardware and a foundation model at the same time. But it means the correct response to the funding news is a calendar reminder, not an evaluation. Compare Meshy, where the individual tiers run from $0 to $100 a month [4] and the free plan gives you 100 credits a month with no credit card required [5]. Both companies sit in the same category of coverage. Only one of them is a purchase decision this quarter.

Portability is the only insurance you can buy in advance

You cannot verify a vendor’s balance sheet and you cannot predict a pivot. The one thing you can check, before you pay anything, is whether the output leaves in a format that outlives the vendor.

Meshy exports GLB, FBX, OBJ, STL, USDZ and 3MF [7]. That list matters more than it looks. GLB embeds textures in one file, while FBX and OBJ use separate texture files you have to keep together [7], so an asset library that loses track of those companion files opens as untextured grey meshes. STL is for single-colour 3D printing, 3MF supports colour and multi-material, and USDZ is described as the only AR format for the Apple ecosystem, used by AR Quick Look [7]. Before you commit, export one finished asset and open it in the next tool in your chain, not in the vendor’s own preview.

Do the same test on any AI tool in a hot category. If the output only exists inside the vendor’s viewer, inside a proprietary project file, or behind an API that returns a URL rather than a file, you are not buying a tool. You are renting a dependency, and the exit cost is the entire body of work.

The licence on your tier is the one that binds you

This is where free tiers quietly break client work. Meshy’s free plan gives 100 credits a month with no card required [5]. Read the terms and the arrangement is more specific than “free”: for customers on the free plan, the provider owns all right, title and interest in the assets, and grants such customer a licence under the Creative Commons Attribution 4.0 International Licence [6]. The documentation states what that licence costs you in practice, listing free-tier commercial use as permitted with attribution [5]. If you hand a free-tier asset to a client inside a game build or a product render without carrying that attribution, you have shipped a licence problem on someone else’s behalf.

Paid tiers are different. All paid plans from Pro upward include commercial-use rights, and the documentation lists the Pro licence type as private, with the user owning the output [5]. The terms add that paid customers have the option to keep their user content private and that it will not be used for any purpose other than as outlined in the agreement [6]. Note that the two pages do not entirely agree: the documentation’s table records the user as owning the output on every tier including free [5], while the terms assign ownership of free-tier output to the provider [6]. When the documentation and the terms of use disagree, the terms are the document you agreed to.

Two more clauses are worth knowing before you subscribe. Fees are non-cancelable and non-refundable unless an order expressly says otherwise [6]. And the provider may, at its sole discretion, delete output associated with an inactive account [6]. None of that is unusual. All of it is invisible from the funding headline, and all of it is decided by which row of the pricing table you are actually on.

Price the work per usable asset, not per month

The monthly price is the least informative number on the page, because AI tools bill in credits and you throw most generations away.

Meshy Pro is $20 a month or $240 a year and includes 1,000 credits a month [4]. Premium is $40, Ultra is $100, and the Studio team plan is $70 with one member included plus $10 a month per additional member [4]. On both Meshy 6 and Meshy 7, a text-to-3D mesh costs 20 credits and AI texturing costs 10, so a full generation runs 30; image-to-3D costs 25 credits on Meshy 7 and 20 on Meshy 6, and remesh, rigging and animation cost nothing [5]. Monthly credits reset at the start of each month rather than banking, though purchased credits do not expire [5].

So 1,000 credits is about 33 full generations, or roughly $0.60 each on Pro. That is the number people quote. It is not the number you pay. If three generations in ten are good enough to keep, your real cost is closer to $2 per finished asset, and a bad prompt-to-keep ratio moves that faster than any plan change. Work out your own keep rate after a week on the free tier, then decide which tier you need. The vendor’s raise has no effect on this arithmetic, which is the point.

What a funded vendor owes you in writing

The last check is the one that separates a company you can build on from one you can only try. Ask what the vendor has committed to in writing about taking things away.

Anthropic publishes a model lifecycle with four defined states, active, legacy, deprecated and retired, and commits to at least 60 days’ notice before model retirement for publicly released models, with notification by email and in the documentation [8]. It also publishes the dated record: claude-opus-4-1 was deprecated on 5 June 2026 and retired on 5 August 2026, and claude-3-haiku was deprecated on 19 February 2026 and retired on 20 April 2026 [8]. Anthropic further states that it has committed to long-term preservation of model weights [8]. Whatever you think of any individual retirement, that is a promise you can plan a quarter against.

Now compare that with the typical newly funded vendor, where the equivalent commitment is a line like the one in Meshy’s terms saying the features included in each plan may be modified or updated from time to time [6]. That is normal and it is also the weaker promise: it tells you the plan you bought can change and does not tell you how much warning you get. A company that has just raised hundreds of millions has more reason than ever to reshape its tiers. The written notice policy, not the bank balance, is what determines whether that reshaping arrives as a scheduled migration or as a broken Tuesday.

What still goes wrong

These checks tell you about terms and portability. They tell you nothing about quality. A vendor can have a clean licence, six export formats and a fair credit price, and still produce meshes with topology you would have to rebuild by hand before a client sees them. The only test for that is your own material through your own pipeline, which is why the free tier exists and why a week on it is worth more than any amount of reading. Run the licence check first, though, so the week on the free tier does not end with something attributable inside a paid deliverable.

The terms you read today are a snapshot. Meshy’s own terms reserve the right to modify plan features from time to time [6], and every vendor in this category has similar language. A large round makes tier changes more likely, not less, because the company now has a reason to move upmarket and the cheap self-serve tier is the one most likely to be repriced. There is no public signal that tells you when that is coming. The only defence is the one you can build in advance: keep exports current, and do not put anything in a workflow that you could not rebuild elsewhere inside a week.

Finally, this framing does not scale down to zero risk, and it is not the right framing for everyone. If you are evaluating a robotics foundation model for a production line, or anything with a hardware dependency and a deployment date, the questions are different and harder, and they belong with people who can run a pilot and read a contract. This guide covers the case where the downside is a wasted subscription and a week of rework, not a stalled factory.