AI cost projection: Anthropic versus own hardware

How to read this. Anthropic only is the full bill from 6 October 2026: metered consumption plus 1,749 seats. Break-even (5% / 5%) breaks even over the two years combined; year 1 is still a loss and year 2 earns it back. Year 2 carries no adoption ramp.

The two growth dials. Token growth drives consumption, so both the metered bill and the hardware. It defaults to 7.3% a month, derived from the vendor's reported spend growth of +36.9% divided by its active-user growth of +27.6%, so it is already a per-user figure. Seat growth moves the seat licence line only. Its 3% default is a planning assumption, not a measurement: billed seats went 1,715 in June to 1,716 in July, so the rollout has flattened. While seats stay ticked they sit on both sides and the saving is unchanged; untick and the dial starts to matter.

Where the numbers come from. July 2026 is anchored on the measured invoice and seat counts per tier. Token volumes are estimated, because there are no per-product token counters and the last telemetry export is dated 21 May 2026: code 177.9 B tokens at €0.60 per Mtok, chat 23.5 B at €4.44, with 88% cache reuse on code and 65% on chat. One measurement window is a thin base for two years, so year 2 is indicative.

The two-model split. Developers get two endpoints: GLM on the B300 nodes for orchestration and planning, Qwen on RTX6000 pairs as subagents. That is not context-length routing, so prefix caching stays intact. The 65% default is a design assumption, not a measurement, since today's telemetry comes from a single-agent world. What supports it: a turn reads only 17,824 fresh tokens and writes 680, and 117 of 206 code developers already use agent patterns. Note the limit of this dial: it decides how much hardware you buy, not what the workload costs. Hardware and run cost are about 23% of the hybrid total; the other 77% is the Claude volume that has not moved, driven by the sliders at the top.

Hardware. B300 node €24,343 a month, RTX6000 pair €2,122. A pair is set at 30,000 prefill tokens per second against 123,000 for a node, holding both tiers to the same 8.2% of peak FP4 that the node figure implies. Published llama.cpp numbers are far lower (7,640 at short context) but are a floor, not a serving estimate: Q4_K never reaches the FP4 cores and at batch 512 the expert GEMMs are too small to be compute-bound. Benchmark this on your own prompt mix before ordering, because it sets the pair count. The two tiers are sized on different peak factors, both adjustable in the settings, and deliberately so: 2.5x for the nodes, 1.5x for the pairs. An orchestrator turn has a developer waiting on it, so it has to be sized for the worst second of the day, but subagent calls are dispatched by the harness and can queue for a few seconds without anyone noticing. Sizing queueable work on the same peak as interactive work is what made the pair count run away. At 1.5x both tiers finish the projection at a comparable load, 99% for the nodes and about 91% for the pairs. The caveat runs the other way on caching: the 88% hit rate is measured on long interactive sessions, and short subagent tasks amortise a prefix cache less well, so if the Qwen tier lands nearer 75% its fresh load roughly doubles and the pair count returns to where it was. Pairs also carry no ops cost in this model, which is generous. Five pairs is 10 cards, not 5 chassis, and the tier carries about as many fresh tokens as the nodes do, so comparable GPU counts on both sides are expected. All hardware prices assume a three-year commitment, including nodes added later. A node is bought only once the overflow to Anthropic costs more than the node itself, which is why the transition is nearly smooth. Capacity is bound by KV memory (208 developers x 6.86 GB plus 400 GB of weights against 2,304 GB) and by peak prefill throughput. Seats are 1,749 x €14.74; left on they keep running, which is the contractual reality until the commit is renegotiated.

The private-AI run cost, and why it is a checkbox. Three layers a metered API bills into its own price. Deployment and inference management, the serving stack itself, scales with nodes: €5,000 a month for the first, €2,500 for each one after. Gateway governance, the routing and policy layer, 0.25 FTE or €2,500. Open-source harness management, the OpenCode-style client the developers actually use, 0.5 FTE or €5,000; this is the layer most likely to be light, since 208 developers depend on it and upstream moves fast. Together €12,500 a month at one node, €17,500 at three, and €322,500 over the two years in the Optimistic case. All estimates rather than quotations, at a blended €10,000 per FTE per month.

What is not modelled. Capability. Every figure assumes the open-weight model is an acceptable substitute on the real workload, which a bake-off still has to establish. Chat is the weaker case: a third of it runs on Opus today. Above roughly 10% total growth a month the arithmetic still holds but the plan does not, because lead time and staffing bind before capacity does.

How to read this. Anthropic only is the full bill from 6 October 2026: metered consumption plus 1,988 seats at $17. Own hardware goes live in December 2026, not October: delivery is realistically December, so October and November are Anthropic on both lines (the start month is a setting). The one RTX 6000 pair that already runs today is on the hybrid line from October; what it serves in the pilot before December is not counted as a saving. It carries today's pilot comfortably, but the plan needs a second pair once the node is live in December and four to five by February 2027 in the Optimistic case. Break-even (3% / 3%) breaks even over the two years combined. Year 2 carries no adoption ramp.

What changed since 13 August. Four corrections, all from measurements that did not exist then. (1) Chat volume was undercounted six to nine times; the dollar totals were right, the token counts were not. (2) Cache reads are 95.6% of all tokens, not 88%. (3) Growth is far faster: +41% a month on average from May to September and +31% in September, against 7.3% before. (4) New input, the part that costs GPU compute, barely grows. Together these make the Anthropic line much steeper and the hardware much cheaper to scale.

The growth dial. Both growth dials start from a single comparison: May 2026 against 6 and 7 October only. Everything in between is left out, because it carries the licence burst before the metered model started on 6 October. Weekday against weekday (Analytics chat, organisation level): new input (uncached plus cache creation) went from 0.52 B to 1.42 B a day, about +24% a month over 4.7 months; total tokens from 12.8 B to 32.2 B a day, about +22% a month. As monthly totals, with 22 weekdays and 9 weekend days in October and the May weekend ratio (26% new input, 35% total), October comes to about 34.5 B new input and 810 B tokens, against 11.1 B and 289 B in May. The two rates are close, so the context per session barely grows. Both are a two-day sample, and the same two days are 25% (new input) and 52% (cache reads) below 29 and 30 September. A constant 22% for two years would multiply volume by 100 or more, so growth holds at the slider value for the first three months (the decay starts after slider), an assumption for how long adoption keeps growing before it peaks, and then decays toward a long-run floor (3% a month, half-life four months, both settings); the decay now applies to new input as well as to total tokens. Untick it to hold the rates constant. Seat growth defaults to 2%: billed seats went from 1,716 in June to 1,969 in September (+4.7% a month), but with seats nearly full that pace cannot hold. While seats stay ticked they sit on both sides and do not move the saving. The October starting point was rebased twice as measurements came in: $200K of metered spend (EUR 173,386) on the 6-7 October two-day sample, then $250K (EUR 216,733) on 15 October, when the first nine full token-based days were in: 6-14 October measured $80.0K, $8.9K a day, weekdays $10.8K and weekend days $2.1K, weekday/weekend-adjusted to $250K per 30 days ($264K if the partial 6 October is excluded). Plus seats, with 34.5 B of new input.

Where the numbers come from. The Anthropic usage report of 15 September 2026 and the GenAI overview of the same day. October is the measured central scenario: the September run-rate of 1,160 B tokens a month grown by 31%, metered at a blended $0.293 per Mtok ($0.5554 at list price x 0.528 enterprise discount, on the September model mix), $445,293 a month or €386,038. Code and chat are split 55% / 45% of that spend, by applying the July per-product rates ($0.26 per Mtok code, $0.59 chat) to the August token mix (73% code, including Cowork and Office add-ins). Seats are 1,988 x $17 = €29,299. Opus is 74% of the spend on 55% of the tokens. The October metered base used by the model was rebased to $200,000 (€173,386) on the 6-7 October sample and then to $250,000 (€216,733) on 15 October, when the first nine full token-based days were in (6-14 October, $80.0K after discount), with the same 55% / 45% split. Fresh input and KV stay on the measured September load, because hardware has to carry the real tokens, not the bill assumption. The $445,293 above is the earlier measured-growth scenario and is no longer the model's cost starting point.

The lesson that changes the hardware: fresh input is flat. In August the account read 855.7 B cached tokens, wrote 35.5 B new input and 4.0 B output. New input went 34.7, 35.5, 37.3 B from July to September (+3.7% a month) while the total tripled. All the growth is context being re-read. Prefill compute follows new input; the re-read context costs KV memory. The model therefore splits the two: prefill load grows with the new input slider (new-input growth), and the KV memory the workload holds grows with the rest of the token growth. From 6 October the slider defaults to 24%, the May to 6-7 October rate (see the growth dial); the 6.6% (growth in active users) and the 3.7% (July to September, a flat stretch) are the slower alternatives. This dial sets the hardware count. At 3.7% one node lasts the full two years; at 6.6% the second node arrives in April 2028; at 10% there are four (measured on the earlier constant-rate model). The saving barely moves across that range, because hardware is small next to the Anthropic bill. On the Qwen pool the same picture holds in production: 1 B prompt tokens on 28 September at a 91.4% prefix-cache hit, 106k tokens per request, and at most 11 requests computing at once.

Hardware. A B300 node now carries two models, GLM-5.3 on six GPUs as orchestrator and GLM-5.3-Flash on two as the Sonnet-class coder, both NVFP4 with FP8 KV. That gives about 540 resident sessions at 148k context (190 + 350) and about 160,000 prefill tokens per second (92,000 + 68,000), for the 265 Claude Code developers the September figures imply. The RTX 6000 pair is now measured, not derived: 12,000 prompt tokens per second at the comfortable limit of the 27 September load test (11,793 at 32 concurrent agents), well below the 30,000 the August version assumed. Its KV pool is 1.87 M tokens per card on BF16 and about 1.7 times that on FP8, sized on the 106k context seen in production with one resident session per Qwen developer. The pair count is higher than in August (12 against 5 in the Optimistic case) for two measured reasons: a pair does 12,000 tokens per second, not 30,000, and production sessions are long. What FP8 KV buys depends on the new-input dial: at 6.6% the pairs run out of prefill before they run out of memory, and FP8 and BF16 both need 12; at 3.7% memory binds first and FP8 needs 8 pairs against 14 on BF16 (switch it in the settings). FP8 stays the right setting, it is the margin for longer sessions. The node was already FP8 in the August version. Peak factors are unchanged at 2.5x and 1.5x over 200 working hours; production showed an hourly peak only 1.4 times the busy-hour average, but an hour is not the worst second. All hardware prices assume a three-year commitment. A node is bought only once the overflow to Anthropic costs more than the node itself.

The cache-discount switch. 95.6% of all tokens are cache reads, billed today at a tenth of the input price. Under the consumption model cost follows token volumes "at the rates on your order form"; whether cache reads keep their discount there is not confirmed in any mail seen so far. The switch is on by default, which is the reading of "cost-neutral on current usage": without the discount the same usage would cost seven times more, which is not cost-neutral. Untick it to see the risk: on the September model mix the blended price goes from $0.5554 to $3.8801 per Mtok at list, so code and chat on the Anthropic line are multiplied by 7.0. Seats and own hardware do not change, because on own hardware a cache read costs memory, not compute.

The private-AI run cost. Unchanged from the August version: deployment and inference management €5,000 a month for the first node and €2,500 for each one after, gateway governance €2,500, open-source harness management €5,000, at a blended €10,000 per FTE per month.

What is not modelled. Capability: the GLM-5.3 and Flash bake-off against Opus and Sonnet still has to run, and chat is the weaker case since Opus carries most of the spend. The migration credits: the agreement signed on 29 September is described as cost-neutral on current usage, with migration credits and discounted usage rates, but the credit amount and the rates on the order form are not public to this model, so year 1 on the Anthropic line is likely too high by the size of the credit. The chain between gateway and vLLM breaks with HTTP 502 above about 192 agentic users, and prefix-cache reuse on the Qwen tier needs sticky routing, which Bifrost does not do today (half of follow-up turns land on the other machine). Decode is not a constraint anywhere in this workload (75 to 218 input tokens per output token).