Equinix has given you a date and withheld a price. On 2 September it announced Equinix Inference Exchange — managed inference for more than 200 open-source models, running on NVIDIA reference architectures inside Equinix data centres, operated by Together AI, available starting Q1 2027. No pricing was disclosed. None will be for months.
That combination — a firm quarter, an empty price line, four to seven months of lead time — lands in the middle of the Q4 renewal season, and it is the wrong thing to wait for and the wrong thing to ignore. The decision it forces is not which venue wins. It is how long a term you sign this quarter, so that when a real quote finally exists you are in a position to make somebody compete for the work.
What Equinix Actually Announced on 2 September
Equinix announced a distribution deal, not a product you can buy today. The press release is precise about the parts: NVIDIA supplies validated Enterprise Reference Architectures, Together AI supplies the inference platform and its catalogue of 200-plus open models, and Equinix supplies the floor space, power, cooling and the interconnect — 280-plus data centres across 77 metros, 230 cloud on-ramps, more than 10,500 interconnected businesses. Three named use cases: metro edge inference, open model migration, and sovereign AI.
The tenancy detail is the one that matters commercially. Both multitenant and dedicated single-tenant deployments are on the roadmap — "multitenant deployments for shared efficiency and dedicated single-tenant environments" for workloads that need their own capacity. Those are two different products with two different cost structures and two different compliance stories, and neither has a number attached to it.
Note also what shipped alongside it. Equinix used the same event to announce Equinix Fabric One, a managed connectivity layer with AWS and Google Cloud as lead integration partners, which "is expected to enter beta later this year, with general availability planned for 2027, initially in North America." Two flagship AI announcements on one day. Two 2027 dates. Zero prices.
Your Token-API Commitments Probably Already Clear March
Here is the good news nobody is telling you: if your AI spend runs through a managed model API, the longest term any of the three hyperscalers will even sell you is shorter than the wait.
- Amazon Bedrock Provisioned Throughput offers exactly three levels: no commitment, one month, or six months. AWS is blunt about what a commitment means — "You can't delete the Provisioned Throughput until the six month commitment term is over." A six-month PT bought today reaches the end of its commitment term in March 2027.
- Azure AI Foundry provisioned throughput sells reservations for a one-month or one-year term. A one-year PTU reservation signed this week runs to September 2027 — past the date, but only by two quarters, and Microsoft's own documentation notes reservations are "cancelable, with limits" and that exchanges "reset the reservation term."
- Google Vertex AI Provisioned Throughput went the other way entirely and added one-week terms for select models, explicitly "without a monthly or yearly commitment."
Read those three together and the direction is consistent: the managed-inference market is selling shorter commitments, not longer ones. None of the three says why. If your exposure is a Bedrock PT or a Vertex GSU order, Equinix's Q1 2027 date costs you nothing. Do not renegotiate anything on account of it.
One caveat that catches people: a reservation is a billing construct, not a capacity guarantee. Microsoft states it plainly — "Purchasing a reservation doesn't reserve capacity on the service" — and tells customers to create the deployment first, then buy the reservation. Shortening a term does not shorten your queue for silicon.
A second caveat costs more, because it fires when you do nothing. On both clouds the default at the end of a term is renewal, not expiry: AWS states that "Your Provisioned Throughput will automatically renew at the end of each commitment term", and Azure's instruction for a reservation you no longer want is to "turn off auto-renew to prevent it from renewing at the end of its term". A term that clears March 2027 only buys you the option if somebody cancels it before it rolls. Put the cancellation date in the sheet, not the end date.
The Deals That Do Lock Past March Are the Ones This Competes With
Equinix Inference Exchange is not aimed at your Bedrock line item. It is aimed at the deal you sign when you decide to stop renting tokens and start renting — or owning — machines. Those are the contracts with real duration.
CoreWeave's own disclosure is the cleanest public number. In the 8-K announcing a $2.6 billion delayed draw term loan facility on 10 August 2026, the company describes a facility with a "five-year maturity while its underlying customer contracts average approximately three years." Three years is the operating assumption in CoreWeave's own book — though read the rest of that release before you treat it as fixed, because co-founder Brannin McBee uses it to argue the opposite direction: "Lenders are now comfortable financing shorter-dated contracts, which allows us to target a wider variety of customers." We have written before about how the neocloud tiers actually price against each other. A three-year capacity contract signed in September 2026 does not surface again for competitive tension until September 2029. Inference Exchange will have shipped, matured, and possibly been retired inside that window.
The same is true of the colocation contract itself, which is where the electricity, the cross-connects and the multi-year renewal clauses live — and where the operating cost is exposed to policy you do not control, as North Carolina's data centre electricity tax fight demonstrated this summer.
There is a useful price anchor available while you wait. Together AI already publishes list rates for the exact service Equinix intends to resell: as of September 2026, its pricing page shows dedicated NVIDIA HGX H100 inference at $5.49 per GPU-hour (discounted to $3.99 through 30 September 2026), HGX B200 at $8.99, and on-demand GPU clusters at $3.99 for H100 and $8.19 for B200, with reserved cluster discounts banded at 7-30, 31-90, 91-180 and 181-plus days. Reserved pricing for the dedicated inference product itself is not published at all — that line reads "Contact sales." That is your reference point for whatever an Equinix-fronted quote turns out to be, and it means the bundle has to justify a premium over a price the partner already publishes.
Equinix's Recent Record on Dated AI Products Is Mixed
Steel-manning this: Equinix is not a startup announcing vapour. It reported $2.625 billion of revenue in Q2 2026, up 16% year over year, $424 million of annualised gross bookings, a record 9,700 net interconnections, and raised 2026 capex guidance to $5.0-6.0 billion with 52 construction projects underway across 33 markets. It has the balance sheet and the floor space, and it has hosted NVIDIA hardware for years — Equinix Private AI with NVIDIA DGX is described by NVIDIA as "a turnkey, ready-to-run AI development platform, hosted and managed by Equinix." The capability is not in question.
The delivery record on announced-and-dated AI software is the part to weigh. On 15 April 2026 Equinix launched Fabric Intelligence, and the availability line in its own release reads: "Fabric Intelligence is available now to preview. To get more information and request access, please register your interest here." We covered it at the time as a bet that network operations is the AI bottleneck. Four and a half months later, the connectivity story is Fabric One, in beta later this year and GA in 2027.
And there is the harder precedent. Equinix retired Equinix Metal, its bare-metal compute service, with the platform sunsetting in June 2026 — roughly nine weeks before this announcement. Network World reported CFO Keith Taylor putting Metal at just 1.25% of company revenue, with one analyst noting bare metal is "a much different business model than pure colocation with interconnection."
That is not a prediction that Inference Exchange fails. It is the reason your contract needs an option rather than a plan. Equinix has recently demonstrated both that it will invest heavily in AI infrastructure and that it will exit a compute service that does not clear its revenue bar.
"Sovereign" Is Doing Real Work in That Press Release
Placement is necessary for a residency obligation. It is not sufficient, and the multitenant SKU is where that distinction bites.
Equinix's framing is that the offering lets AI workloads "run in locations that support data residency and sovereignty requirements". True as far as it goes — a metro is a place, and a cage in Frankfurt is in Frankfurt. But the operator of the inference platform is a US company, and Together AI's own privacy policy states that information "may be transferred to, and maintained on, computers located outside of your state, province, country, or other governmental jurisdiction where the data protection laws may differ from those in your jurisdiction," and treats Zero Data Retention as a setting a customer enables rather than a default.
Together's documentation on dedicated endpoints describes them as serving "a model on reserved hardware," billed "per minute by hardware while a deployment runs, regardless of model or request volume." That is a statement about performance isolation and billing. It is not, on its own, a statement about which legal entity processes your prompts, where the control plane runs, or where telemetry and logs land.
If your residency requirement is contractual — a regulator, a customer commitment, a data protection agreement — the answer will come from the DPA and the subprocessor list, not from the metro. Assume for planning purposes that only the dedicated single-tenant configuration can carry that weight, and price accordingly. This is the same trap we mapped when Mistral asked for multi-year European Compute Unit money and its regional endpoint quietly dropped capabilities.
Three Moves Before You Sign Anything This Quarter
This Week: Pull every AI compute and inference commitment into one sheet with three columns — vendor, end date, and what it costs to exit. Include the token APIs, the GPU reservations, the colocation MSAs and the model-vendor capacity commitments. Most teams cannot answer "what expires before April 2027?" in under a day, and that question is now the whole decision. Flag anything ending after Q3 2027.
This Month: For any new capacity commitment over 12 months, negotiate one of two things and do not sign without at least one. Either cap the term at 12 months with a unilateral renewal option at the same rate, or take the longer term and attach a re-price trigger: a clause letting you re-open pricing once if a materially comparable managed offering becomes generally available in your metro. Vendors resist term caps because they price on duration; they resist re-price triggers less, because the trigger only fires if a competitor actually ships. Ask for the trigger first.
This Month: Build the comparison you will need in Q1 2027 now, while nobody is selling to you. Take your top two production workloads, measure real tokens per second at your real context length — the batch-size arithmetic that makes vendor throughput numbers misleading applies here — and price them three ways: current API, self-hosted on a vLLM-class runtime, and Together's published dedicated rate as the colo proxy. A quote you cannot evaluate in a week is a quote you will accept on the vendor's timeline.
Before Renewal: Ask Equinix, in writing, for three things: an indicative price range for both tenancy modes, the metros in the Q1 2027 launch set, and the subprocessor list for the multitenant service. You will probably not get all three. Which ones you do get tells you how real the date is, and it costs you one email.
The Bottom Line
We have been here before, with roughly this shape. In 2010 the question was whether to sign a three-year managed hosting deal or wait for the cloud to be enterprise-ready; the people who did best were not the ones who guessed right about the venue, they were the ones who refused to buy duration they did not need. The same discipline applied to OpenAI's multi-year capacity offers earlier this year, and it applies now.
A dated, unpriced product is not a reason to wait and it is not a reason to hurry. It is a reason to make sure the deal you sign in the next ninety days is one you can re-open in eighteen months without paying for the privilege. Equinix has told you when it wants to sell you something. Structure your contracts so that you are still a buyer when it does.
Buy the option, not the opinion.
Continue Reading
Why Network Ops Is Your Next AI Bottleneck GPU Clouds Compared: Nebius on Price, CoreWeave at 3AM Mistral Wants Multi-Year Money. Its EU Endpoint Drops Agents. North Carolina Taxed the Power. Inference Pays. vLLM vs TensorRT-LLM vs SGLang: Default to vLLM OpenAI's $30M Capacity Trap: When 3-Year Lock-In Pays