NVIDIA GTC Berlin runs from 20 to 22 October at STATION-Berlin, with the keynote at Tempodrom on the Wednesday. We are there this week. No booth, no stand, no branded tote bags — we are a small team and this is a listening trip.
What we would actually like to talk about
Running a router across thirty-odd inference providers gives you one unglamorous dataset that almost nobody else collects: which providers are actually serving, right now, as opposed to which ones have a status page that says they are.
We built a synthetic probe for this a few weeks ago — one token, every configured provider, once a day — because our existing alerting could only see failures on traffic we were already sending. A provider nobody happens to be calling generates no errors, so it can be dead indefinitely and look fine.
The first run was not comfortable reading. Eight providers could not serve at all:
- •Four out of credit, returning 402 with a balance message.
- •Three on credentials that had quietly stopped authenticating.
- •One returning HTTP 410 — it had retired its API outright and we were still carrying its models.
None of that had alerted. Nothing was on fire, because nothing was being asked of them. If you run inference across multiple vendors and you only find out a provider is down when a user complains, that is the same blind spot, and it is worth half an hour of your time to go and check.
The other thing the data keeps showing
Identical weights cost wildly different amounts depending on who is serving them. When NVIDIA released Nemotron 3.5 Super VL last week, eight providers in our catalogue had it within days — at $0.0001 per 1K input tokens on the cheapest and $0.004 on the dearest. Forty times, for the same model.
That is not anybody behaving badly. Providers differ on hardware, batching, quantisation, how much spare capacity they hold and how much they want the traffic. But it does mean “what does this model cost” is not a question with one answer, and a benchmark table that quotes a single price per model is quietly comparing one vendor's business decision against another's.
If you are at GTC with a serving stack and a price, we are interested in the part of that number which is hardware and the part which is strategy. Genuinely — that distinction is hard to see from the outside and we have no way to measure it.
What we are not doing
We are not launching anything this week, we have nothing to demo on a stand, and there is no pipeline to get into. Antbase routes requests to the best model for each one across a few hundred thousand catalogue rows; if that is useful to you it is useful whether or not we meet in Berlin.
The honest reason to go is that an inference-routing layer is downstream of everything being announced in that building. Hardware decisions become serving economics, serving economics become the price spread above, and the price spread is the thing our router exists to navigate. It is worth understanding the top of that chain rather than only the bottom.
Say hello
If you are there and any of this is your problem too — multi-provider reliability, pricing that moves under you, agents that need somewhere safe to run code — come and find us, or reach us afterwards. We will write up anything worth sharing.
