What Does Together AI Do? The Independently-Capitalized Inference Model

Together AI sells open-model inference and GPU hours; investors capitalize more than 500 MW of compute separately from the $800 million equity — cash sticks when tokens or reserved hours bill, while those megawatts are still a clock.

NEOCLOUDS · Data as at 27 Oct 2026 · Next update: after a dated live-megawatt or GAAP-revenue reprint
Affects: Neoclouds · US · Hyperscalers

It is a neocloud: specialist GPU time, not a full hyperscaler platform. The cash job is tokens and GPU hours on the same fleet, with new halls funded off the equity round. See GPU as a service, provisioned throughput, and who finances GPU clusters.

In short: On 1 July 2026 Together printed an $800 million Series C at $8.3 billion post-money, $1.15 billion of annual bookings, and more than 500 MW of compute to be capitalized independently by new investors. IBM, on 11 August 2026, printed a $240 million multi-year HGX B300 cluster on IBM Cloud, expected Q1 2027, and said Together reports 400 trillion tokens a month. Together gets paid when a token or a GPU-hour bills. Live megawatts, not the bookings line, are the stuck step.

Education only, not advice to buy or sell any security. Together AI is private. Figures are from company and IBM releases and live company pages dated as shown. Outlook is labelled outlook. Secondhand reports are labelled reported.

Numbers that matter

Sources: Together AI Series C blog, 1 Jul 2026; Business Wire, 1 Jul 2026; IBM newsroom, 11 Aug 2026; Together AI pricing, checked 27 Oct 2026. Status chips are Second Order’s read.

Our arithmetic. Bookings of $1.15 billion are 1.44× the $800 million equity round (1,150 ÷ 800 = 1.4375). The round is 0.70 years of that run-rate (800 ÷ 1,150 = 0.696), about 8.3 months. Reserved H100 at $3.19 versus on-demand $3.99 is 20% cheaper ((3.99 − 3.19) ÷ 3.99 = 0.201); reserved B200 at $6.79 versus $8.19 is 17%. Dedicated H100 at $5.49 versus the cluster $3.99 is a 38% premium (5.49 ÷ 3.99 = 1.376). IBM’s $240 million is 21% of one year of bookings (240 ÷ 1,150 = 0.209) — only if Together is the payer, which IBM’s wording supports but does not itemise. None of this is utilisation, and none of it prices 400 trillion tokens. There is no public token list.

What does it actually do?

Three steps, in order.

Serve. Fund off-book. Bill two meters.

1. It serves open models as production inference. The 1 July blog names Cognition, Decagon, Eleven Labs, Cursor and Suno, and says Decagon cut inference costs sixfold. Business Wire adds $1.15 billion of annual bookings last quarter. Together is the inference layer. See AI tokens vs GPU hours.
2. It funds new megawatts beside the equity, not from it. “In addition to the equity capital, we have secured commitments for over 500 MW of compute capacity to be capitalized independently by our new investors” (Together blog, 1 Jul 2026). The $800 million is for the platform; the halls are a separate capital stack. IBM’s named hall is a $240 million multi-year HGX B300 cluster on IBM Cloud, expected Q1 2027. Together’s 6 October blog calls itself the first customer of that cluster — a customer claim, not a live-megawatt print. See data-center financing and GPU financing.
3. It bills two meters on the same silicon. Serverless and dedicated endpoints meter the model. GPU clusters meter the card — on-demand, or reserved up to six months paid upfront (GPU Clusters page, checked 27 Oct 2026). Dedicated inference bills the replica while it is ready. That is provisioned throughput. See reserved vs on-demand GPUs.

How does it make money?

It gets paid when a customer uses a working model or a working GPU and pays for that unit. Three things decide whether the model clears:

• Which meter is on the invoice — a token, a dedicated replica-hour, or a cluster GPU-hour. See GPU as a service.
• Utilisation — a serverless GPU that sits idle costs Together; a dedicated replica that sits idle still bills the customer for the hour. See GPU utilization.
• Live megawatts — bookings of $1.15 billion and 500 MW independently capitalized are not Q1 2027 IBM hours, and they are not 400 trillion billed tokens.

There is no public income statement. The dated cash prints are company bookings ($1.15 billion, 1 Jul 2026) and IBM’s $240 million cluster agreement (11 Aug 2026). Bookings are not GAAP revenue and not CoreWeave’s 10-Q backlog. See TCV and take-or-pay GPU contracts.

Who gets paid. NVIDIA, when trays ship into independently capitalized megawatts and the IBM cluster. IBM Cloud, on the $240 million deploy. Investors who capitalize the 500 MW, as halls are built. Together, when a token or an hour bills. Who waits: anyone reading $1.15 billion of bookings, 500 MW, 2 GW+ of “portfolio” capacity, or a 50-fold five-year outlook as GPU revenue today. Portfolio is not live. Outlook is not a clock.

Money sticks at billed tokens and billed hours on machines that are on. CoreWeave’s Q2 2026 10-Q is the contrast: about $104 billion backlog and about $35.1 billion of debt. Together’s disclosed clock is equity plus investor-capitalized megawatts, a six-month prepaid reserve, and an inference token print with no public price. See What Does CoreWeave Do?.

What limits it?

Live halls first. Conversion second. Mix third.

1. Independently capitalized is not energised. The 1 July blog’s 500 MW is a commitment for investors to capitalize. The GPU Clusters page, checked 27 October 2026, also prints “2GW+ in the portfolio with 600MW of near-term capacity in US.” Product-page lines, not a live-megawatt filing. IBM’s Q1 2027 print is the only dated hall with a counterparty and a dollar figure. Until handover, the $1.15 billion bookings line cannot all be served on new silicon. That is the stuck step.
2. 400 trillion tokens a month is not an invoice. IBM (11 Aug 2026) said Together reports 400 trillion tokens monthly. Together’s 6 October blog said “hundreds of trillions.” Neither print gives a price per million. Do not convert the count into revenue.
3. Dedicated replicas bill idle. Dedicated inference bills by the minute for each ready replica, not per token. Thin traffic still pays the $5.49 H100 hour — or scales to zero and Together holds unused cluster capacity. See provisioned throughput and our Wire on Nebius’s idle-GPU tax.
4. Reserved cash is short. Reserved clusters are “up to 6 months, pay upfront.” That is prepay, not multi-year take-or-pay. Cash can arrive before the hour and still leave utilisation risk when the block ends.
5. There is no public live-megawatt, utilisation, or debt schedule. Colo, power and staff still sit under the hour. Expansion still waits on speed-to-power. The halls live on the US door.

The honest uncertainty is conversion. $1.15 billion of bookings, 400 trillion tokens a month, and 500 MW independently capitalized are dated facts. The split — tokens versus hours, at what utilisation, on how many live megawatts — is not in a filing.

Who else does this?

Roles, not a league table.

What separates them is who paid for the GPU, who holds the idle hour, and whether cash is prepaid, borrowed, or earned only when the token or the hour runs. See neocloud vs hyperscaler and who the neoclouds are.

What would change this view?

The read on the model weakens if:

• A later company print still says “$1.15 billion” bookings with no live-megawatt or GAAP-revenue split
• The IBM cluster slips past Q1 2027 while 400 trillion tokens a month still has no price
• Independently capitalized megawatts stay a commitment and never show a handover date
• On-demand H100 or B200 list falls while dedicated replica-hours stay full — conversion is softening
• Reserved six-month prepays are not renewed when blocks end
• A GPU-backed loan appears on Together’s own paper after the “capitalized independently” claim

It stays intact while bookings are a company line, 500 MW is capitalized by investors rather than printed as live, IBM’s hall is dated Q1 2027, and there is no take-or-pay backlog in a filing.

Practical takeaways

• Model: Serve → Fund off-book → Bill two meters. Tokens and GPU hours. New MW sit with investors.
• 1 Jul 2026: $800m Series C at $8.3bn post-money; $1.15bn annual bookings; >500 MW independently capitalized
• Our arithmetic: bookings are 1.44× the equity round; the round is 0.70 years of that bookings run-rate
• 11 Aug 2026: IBM $240m multi-year HGX B300 cluster, expected Q1 2027; Together reports 400 trillion tokens a month
• List, 27 Oct 2026: H100 $3.99 on-demand / from $3.19 reserved; dedicated H100 $5.49 (38% premium to the cluster hour)
• Stuck step: live megawatts — investor-capitalized 500 MW and IBM Q1 2027 — not the bookings line
• Next hard reads: a dated live-MW or utilisation print; IBM cluster handover; any GPU-backed loan on Together’s own paper; a token price

FAQ

What does Together AI do?
Together AI is a specialist GPU cloud that sells open-model inference and NVIDIA GPU hours. On 1 July 2026 it printed an $800 million Series C at an $8.3 billion post-money valuation, $1.15 billion of annual bookings, and more than 500 MW of compute to be capitalized independently by new investors. It gets paid when tokens or GPU hours bill.

What does Together AI do in simple terms?
It runs other companies’ open AI models on rented NVIDIA machines, and also hires those machines out by the hour. New data halls are meant to be paid for by investors, not out of the $800 million equity round.

Is Together AI profitable?
It has not published a profit or loss. The dated company print is annual bookings of $1.15 billion as of 1 July 2026. That is not net income and not a filing.

How is Together AI different from CoreWeave?
Both sell specialist AI compute. CoreWeave funds GPUs with debt and bills mostly on multi-year take-or-pay after delivery (Q2 2026: $2.58 billion revenue, about $104 billion backlog, about $35.1 billion of debt). Together’s disclosed clock is equity plus investor-capitalized megawatts, a six-month prepaid reserve, and an inference token print. See What Does CoreWeave Do?.

What is Together AI’s backlog?
The company has not printed a GAAP backlog or TCV line in a filing. It prints annual bookings of $1.15 billion (1 July 2026). Bookings are not contracted take-or-pay. Treat 500 MW independently capitalized and 400 trillion tokens a month as capacity and usage, not as CoreWeave-style backlog.

Should I buy Together AI shares?
We do not make buy or sell calls, and nothing here is a recommendation. Together AI is private. What we can tell you is what it does, how it gets paid, and which figures would change the picture. The decision is yours.

Keep reading on the map

• The layer: Neoclouds — signed, live, paid
• The term: Neocloud, explained
• The mechanism: AI tokens vs GPU hours · Provisioned throughput · GPU utilization · GPU as a service
• The contrast: Neocloud vs hyperscaler · Take-or-pay GPU contracts · Reserved vs on-demand · Who finances GPU clusters
• The peers: What Does CoreWeave Do? · What Does Lambda Do? · What Does Lightning AI Do? · What Does Nebius Do?
• The Wire: Nebius’s idle GPU tax · Lambda’s $1bn delayed-draw loan
• Where it happens: US

The members cut on the long clock is already live: Where does the stuck step move?. This page stays the free Stack read of the Together AI model.

How we check this

Figures come from Together’s 1 July 2026 Series C blog and Business Wire release, IBM’s 11 August 2026 newsroom print, Together’s 6 October 2026 IBM Cloud blog, and live pricing pages checked 27 October 2026. The 2 GW+ / 600 MW US lines are product-page claims, not filings. Arithmetic is shown in full. We dropped a ~2,000 GPU count on the IBM cluster that we could not tie to IBM or Together (Reuters, reported, not in the table).

Last reviewed: 27 Oct 2026 · Next review: after a dated live-megawatt, utilisation, or GAAP-revenue reprint, or IBM cluster handover

Spot an error? Tell us; we will correct it here. Editorial standards →

Sources

1. Together AI, Announcing our $800M Series C, 1 Jul 2026 — $800m; >500 MW independently capitalized; named customers; Decagon sixfold
2. Together AI / Business Wire, Raises $800 Million at $8.3 Billion, 1 Jul 2026 — post-money valuation; $1.15bn annual bookings; ~50-fold five-year outlook
3. IBM, multi-year agreement, 11 Aug 2026 — $240m; HGX B300 on IBM Cloud; expected Q1 2027; 400 trillion tokens monthly
4. Together AI, Expanding enterprise inference capacity with IBM Cloud and NVIDIA, 6 Oct 2026 — first customer of the dedicated IBM Cloud inference cluster
5. Together AI, Pricing, checked 27 Oct 2026 — cluster and dedicated list hours
6. Together AI, GPU Clusters, checked 27 Oct 2026 — reserved up to 6 months pay upfront; 2GW+ portfolio / 600MW near-term US (product page)
7. Together AI, Dedicated inference pricing docs, checked 27 Oct 2026 — billed by the minute while a replica is ready
8. CoreWeave, Second Quarter 2026 Results (SEC filing), 11 Aug 2026 — contrast figures on What Does CoreWeave Do?