By Thorsten Meyer | October 26, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The company that sold the world its mainframes now tells enterprises to download the model and run it in their own racks. IBM’s Arvind Krishna says open-weight models are “90% as effective” at a tenth of the token cost. The independent scoreboard says the gap is smaller than it was, but not that small. And while companies do the maths on racks, hobbyists found a hidden speed limit in their old GeForce cards and switched it off.
Key takeaways
- The claim: On IBM’s Q3 earnings call on October 21, Arvind Krishna said open-weight models cost “25% to 10% of the token cost of the frontier models” while being “90% as effective”, and that a large enterprise can start with “two or three racks”. IBM sells on-prem hardware, so it is an interested party.
- The independent check: On the Artificial Analysis Intelligence Index (read October 26), the best downloadable model, Xiaomi’s MiMo-V2.6-Pro, scores 46.3; the best proprietary model, Claude Opus 5.5, scores 57.6. That is 80%, not 90%, up from 51% at the end of 2024.
- The asterisk: Kimi K3.1 Pro scores 55.1, or 96% of the leader, but Moonshot has only promised the weights “by Nov 4”. Until they ship, it is a closed model.
- Cheap, but not automatically: DeepSeek’s hosted V4.1 Flash 1017 (MIT-licensed weights, index 45.1) costs 3% to 6% of Opus 5.5’s list price. But OpenAI’s closed GPT-6 Luna costs $0.10/$0.50 per million tokens. On-prem is a bet on volume and utilization.
- Rent vs. buy, with list prices: An 8-GPU H200 server lists from $172,643. Renting the same GPU class on AWS costs $63.30 an hour. At full load the purchase pays off in about 114 days; at 25% load it takes about 15 months, before staff, cooling and networking.
- The enthusiast angle: An open-source Linux module, gpu-unlock, lifts a tensor-core speed limit in RTX 30 cards; forks extend it to RTX 40 and 20. Matrix benchmarks gain up to 80%, LLM prompt processing 14% to 36%, text generation almost nothing. It exploits a flaw in Nvidia’s boot firmware, and Nvidia’s GeForce licence forbids circumventing “any technical limitation”. Nvidia has not commented.
Products Worth Considering
Anker SOLIX C200 DC Portable Power Station, 192Wh, 200W
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Sony WH-1000XM4C Wireless Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
What Krishna actually said
The quote that made the headlines came in the Q&A, not the prepared remarks. Krishna described a gap that closed in two steps. In late 2024, he said, on a broad benchmark “the top model might score a 70, and the open model might score a 20 or a 30.” In late 2025, “it was 70 to 50.” Then the line that travelled: “Would you like something that’s 10% of the cost if it’s 90% as good as the best, or do you want to pay that 10 times higher cost?”
His examples of work that doesn’t need the best model are deliberately boring: “summarizing an email, examining a contract, answering a customer call, composing an email.” And his reason for on-prem goes beyond price. Banks and traders, he said, want to run AI “in a way that is sovereign”, so that “nobody can take any of the underlying data or underlying logs.” Open-weight models give you “the permission and the IP rights to run them on either on-premise infrastructure or on private clouds.”
Two things in the same answer get less attention. First, he named the bottleneck himself: “Where are my kernel engineers? Where are the people who worry about the underlying network?” He said that gap “is probably still there, except for the most sophisticated enterprise”, and he called the frontier cloud “four times more expensive” but “the easy answer”. Second, he is talking his book. He expects on-prem demand to help sales of IBM’s Spyre cards, and The Register notes that IBM’s mainframe revenue fell 20% in the quarter, 18 months into the z17 cycle. In the prepared remarks IBM also claimed its llm-d project helps run large models “at 5 to 10 times lower cost”. That is IBM’s claim, not a measurement I could check.
What the independent scoreboard says
Krishna’s 70, 50 and 90 are illustrations, not a named benchmark. So I checked them against Artificial Analysis, which runs the same ten evaluations on every model and labels whether its weights are downloadable.
| Date | Best proprietary | Best open weights | Open as share of closed |
|---|---|---|---|
| End of 2024 | o1: 15.2 | Qwen2.5 72B: 7.7 | 51% |
| End of 2025 | GPT-5.2 (xhigh): 30.4 | MiMo-V2-Flash: 22.4 | 74% |
| End of June 2026 | Claude Fable 5 (max): 49.6 | GLM-5.2 (max): 33.7 | 68% |
| October 26, 2026 | Claude Opus 5.5 (max): 57.6 | MiMo-V2.6-Pro: 46.3 | 80% |
| If Kimi K3.1 Pro ships its weights | Claude Opus 5.5 (max): 57.6 | Kimi K3.1 Pro: 55.1 | 96% |
All values from the models in Artificial Analysis’s open-vs-proprietary chart, on the current index version, read October 26, 2026.
So the direction Krishna describes is right: open models went from half to four fifths of the leader in under two years. But in points, the gap has not shrunk. It was 7.5 points at the end of 2024 and is 11.3 today. Closed models moved too. The “90%” is reachable only if Kimi K3.1 Pro’s weights arrive as promised, and as I wrote in my Kimi K3.1 Pro analysis, no licence has been published and a 2.8-trillion-parameter model is a data-centre project in itself.
Licences matter as much as scores. MiMo-V2.6-Pro and DeepSeek V4.1 Flash 1017 ship under MIT; GLM-5.3 (44.8) and Step 5 Preview (43.7) under their own commercial licences. Meta’s Muse Spark Open has conditions, MiniMax M3.1 Flash arrived on Hugging Face this week under a custom community licence, and NVIDIA’s permissively licensed Nemotron 3.5 family sits far lower on the index. Read the licence before the leaderboard.
IBM's 90% claim, checked
Open weights reach 80% of the leader
Krishna says open-weight models are "90% as effective". On an independent index the best model you can download today reaches 80% of the best closed one, up from about half two years ago.
Best open model as a share of the best closed
Artificial Analysis Intelligence Index, current version, best downloadable vs. best proprietary model at each date.
What Krishna said
Illustrations from the earnings call, not a named benchmark.
- Late 2024"70" vs. "20 or 30"Krishna, Q3 call, 21 Oct · no named benchmark
- Late 2025"70 to 50"Krishna, Q3 call, 21 Oct · no named benchmark
- Now"90% as effective"Krishna, Q3 call, 21 Oct · no named benchmark
Top downloadable models today
Intelligence Index score and licence type. Read the licence before the leaderboard.
| Model | Index | Licence |
|---|---|---|
| MiMo-V2.6-Pro Xiaomi | 46.3 | MIT |
| DeepSeek V4.1 Flash 1017 DeepSeek | 45.1 | MIT |
| GLM-5.3 Z AI | 44.8 | own commercial licence |
| Step 5 Preview StepFun | 43.7 | own commercial licence |
| Muse Spark Open Meta | 38.9 | licence with conditions |
In points the gap did not shrink: 7.5 points at the end of 2024, 11.3 today. Closed models moved too.
Sources: Artificial Analysis, Intelligence Index v4.3.2 (read 26 Oct 2026), IBM Q3 2026 earnings call transcript (21 Oct 2026), Artificial Analysis on Kimi K3.1 Pro (23 Oct 2026).
Powered by Thorsten Meyer AI
Where “a tenth of the price” comes from
The price claim holds up best when you rent an open model from its maker. DeepSeek’s official price list for V4.1 Flash 1017 is $0.15 per million input tokens and $0.60 per million output tokens off-peak, double that at peak. Claude Opus 5.5 lists at $4 and $20. That makes DeepSeek 3% to 6% of the price, at an index score of 45.1 against 57.6. My DeepSeek V4.1 Flash 1017 review has the per-task numbers.
But a hosted API is not on-prem. Your prompts and logs then sit on DeepSeek’s servers, which is exactly what Krishna’s bank CEO wants to avoid. And the closed side is not standing still: as I showed in the AI price war piece, OpenAI’s GPT-6 Luna now costs $0.10 per million input tokens and $0.50 per million output tokens, cheaper than DeepSeek at peak, at an index score of 38.1. Meanwhile Muse Spark Open, an open model, costs the same as Meta’s closed Muse Spark 1.3 when you rent it from third parties. Price follows who hosts the model, not whether its weights are public.
The turn: on-prem isn’t free
Here is the arithmetic with list prices, so you can redo it with your own quotes.
- Hardware: BIZON lists an 8-GPU NVIDIA HGX H200 server from $172,643. Eight H200s have 1,128 GB of GPU memory (141 GB each, per NVIDIA), enough for a model of roughly 550 billion parameters at 8-bit, like DeepSeek V4.1 Flash, with room for context. A single workstation card, the 96 GB RTX PRO 6000 Blackwell, now costs $16,000 at NVIDIA’s US store, up from $8,565 at launch (Tom’s Hardware).
- Renting the same class: AWS lists the p5en.48xlarge, with eight H200s, at $63.30 an hour on demand in N. Virginia. The purchase price equals about 2,730 rented hours: roughly 114 days at full load, about 15 months at 25% load, about three years at 10%.
- Energy: NVIDIA rates each H200 at up to 700 W, so eight draw up to 5.6 kW before CPUs, fans and cooling. NVIDIA rates a DGX B300 system at about 14 kW. At the US commercial average of 14.47 cents per kWh (EIA, August 2026), that is roughly $7,100 to $17,700 a year, running around the clock.
- The fixed bill in tokens: Spread over three years, the H200 server plus GPU power is about $65,000 a year. That buys about 3.2 billion output tokens a year from Claude Opus 5.5 at list price, or about 54 billion from DeepSeek’s hosted API at peak rates.
What the list leaves out is what Krishna himself flagged: people. Someone has to choose and quantize models, tune serving software, cluster GPUs, cache results, patch the stack and swap models when a better one ships next month. The security side is not abstract this month: the CVE wave and Broccoli2Shell both hit the kind of open-source plumbing that model servers run on. On-prem means those patches are yours.
Cloud APIs still win when volume is low or spiky, when a task genuinely needs the top model, when you have no team to run inference, or when you want to switch models every few weeks. On-prem wins when data cannot leave, when volume is high and steady, and when latency must be predictable.
On-prem is a volume bet
Rent or buy: the break-even depends on load
At full load an 8-GPU server pays for itself against cloud rent in months. At 10% load it takes years, and people, cooling and patching come on top.
When does buying beat renting?
Purchase price vs. on-demand rent for the same GPU class. Hardware only.
The yearly fixed bill, in tokens
~$65,000 a year: the server over three years plus GPU power. The same money buys this many output tokens.
Before you buy racks: the checklist
- Data: must prompts and logs stay in-house?
- Volume: billions of tokens a year, and steady?
- Quality: which tasks really need the top model?
- Latency: predictable response vs. elastic capacity
- People: who serves, quantizes, patches, swaps models?
- Licence: MIT and Apache are simple; custom ones are not
- TCO: three years of hardware, power, space, support, staff vs. API spend
Our arithmetic from list prices read 26 Oct 2026. Leaves out staff, cooling, networking, space and discounts. GPU power: 8 × 700 W (H200, NVIDIA); 14 kW: NVIDIA DGX B300 rating; electricity: US commercial average 14.47 ¢/kWh (EIA, August 2026).
Sources: BIZON X9000 G3 list price, AWS p5en prices via RunsOn (AWS Pricing API), NVIDIA H200, NVIDIA DGX B300, EIA Electric Power Monthly, DeepSeek pricing, Anthropic pricing, all read 26 Oct 2026.
Powered by Thorsten Meyer AI
The enthusiast twist: a hidden speed limit in GeForce cards
On October 18 a developer called Zagitta published gpu-unlock, a Linux kernel module that “doubles the instruction issuing speed of FP/BF16 FMA into FP32 operations” on RTX 30 cards. In plain words: GeForce cards run one type of tensor-core maths at half the rate of their professional siblings, and the module lifts that cap. It reached Hacker News on October 24 with 345 points and had about 9,200 GitHub stars by October 26.
How it works matters for the risk. The module loads before Nvidia’s driver, writes two speed-selector registers and resets the card, then hands it to the unmodified driver. It is not a flashed BIOS and not a patched driver, and the setting is lost when the card powers off. To get there, the project’s own write-up explains, it uses a buffer overflow in Nvidia’s signed boot firmware, found after Nvidia published a debug firmware image encrypted with a known test key.
The numbers, from the project’s README on an RTX 3090 Ti at 450 W:
| Precision | Before | After | Change |
|---|---|---|---|
| FP16, FP32 accumulate | 87.46 TFLOPS | 118.39 TFLOPS | +35% |
| BF16, FP32 accumulate | 87.50 TFLOPS | 157.41 TFLOPS | +80% |
| TF32 | 42.58 TFLOPS | 82.61 TFLOPS | +94% |
User reports for other RTX 30 cards range from +25% to +84% on BF16 matrix maths; the lowest came from cards capped at 225 W. A fork, gpu-unlock-ada, reports RTX 4090 FP16 instruction throughput going from 165.6 to 331.6 TFLOPS, which is where the “2×” headlines come from. Its own README is more careful: in a matched owner run, real cuBLAS matrix multiplication gained 12% to 34% for FP16, and a quantized mixture-of-experts workload on a modified 48 GB card showed no gain at all. A Turing port reports an RTX 2080 Ti going from 52.8 to 77.9 TFLOPS. These are community measurements on single cards, not independent lab tests.
What it does for local AI: LLM prompt processing (prefill) ran 14% to 36% faster in the reports the README links. Text generation, which is limited by memory bandwidth, and INT8 work stayed about the same. As I explained in the Mac Studio local-AI piece, bandwidth decides how fast tokens come out. Outputs that reports compared were bit-identical.
The risks are real. The Turing port warns that full-speed tensor cores “may damage your hardware. Use at your own risk.” MakeUseOf’s write-up warns that it voids your warranty. Most configurations are tested on one or a few cards. And Nvidia’s GeForce software licence says users may not “bypass, disable, or circumvent any technical limitation” in the software and that GeForce software “is not licensed for datacenter deployment”. Whether a register change made before the driver loads falls under that clause is a question for lawyers; I found no ruling. Nvidia had made no public statement by October 26. Wccftech also reports used RTX 3090 prices jumping from about $1,000 to over $2,000 since the news, a single-source claim I could not verify.
gpu-unlock, explained
GeForce cards run one kind of tensor maths at half the rate of their professional siblings. An open-source Linux module lifts the cap. Big in benchmarks, smaller in real AI work.
RTX 3090 Ti after the unlock
Dense 8192³ GEMM at 450 W, TFLOPS, five-sample medians (project README).
What gets faster in real work
| Workload | Change | |
|---|---|---|
| LLM prompt processing (prefill) RTX 30, user reports in the README | +14 to 36% | faster |
| LLM text generation (decode) limited by memory bandwidth | about 0% | no gain |
| INT8 work not the unlocked path | about 0% | no gain |
| cuBLAS FP16 GEMM, RTX 4090 matched owner run, Ada fork | +12 to 34% | faster |
| Quantized MoE serving, modded 48 GB RTX 4090 one card, one workload | 0.983× | no gain |
Risks before you try it
- Warranty: MakeUseOf warns the unlock voids it
- Hardware: the Turing port warns it “may damage your hardware”
- Licence: Nvidia’s GeForce software licence bans circumventing “any technical limitation” in the software and excludes “datacenter deployment”
- Method: exploits a buffer overflow in Nvidia’s signed boot firmware
- Evidence: community reports, one or a few cards per model
- Nvidia: no public statement as of 26 Oct
The module writes two speed-selector registers before Nvidia’s driver loads and resets the card; nothing is flashed, and the setting is lost at power-off. RTX 40 (Ada) and RTX 20 (Turing) support comes from forks.
Sources: Zagitta/gpu-unlock README and write-up, gpu-unlock-ada README, turing-unlock README, MakeUseOf (26 Oct), NVIDIA GeForce software license, Hacker News, read 26 Oct 2026.
Powered by Thorsten Meyer AI
What this means for companies
Treat Krishna’s pitch as a good question, not an answer. Before you buy racks, run this checklist:
- Data sensitivity: Must prompts, documents and logs stay in your building or your private cloud? If yes, open weights on your own hardware may be the only option, and price becomes secondary.
- Volume and utilization: Estimate tokens per day and how steady they are. Below a few billion output tokens a year, a cheap API (open or closed) usually costs less than a server that sits idle.
- Quality needs per task: Route by task. Email summaries and contract checks may be fine on an 80% model; agentic coding or hard analysis may not. Test on your own data, not on an index.
- Latency: On-prem gives predictable latency and no rate limits. APIs give you elasticity.
- Team skills: Who serves, quantizes, patches and swaps models? If nobody, budget for people or a managed private deployment first.
- Licence: MIT and Apache are simple. Custom and “community” licences may restrict use, users or derivatives. Kimi K3.1 Pro has none yet.
- Total cost of ownership: Hardware, power and cooling, space, networking, support contracts and people over three years, compared with API spend at your real volume, including discounts you can negotiate now.
Do not run unlocked GeForce cards in production. Beyond warranty and stability, the GeForce licence excludes data-centre deployment.
What this means for users and enthusiasts
Local AI on your own GPU is good and getting better. A 24 GB card such as the RTX 3090 holds a model of roughly 27 billion parameters at 4-bit. Alibaba’s Apache-licensed Qwen3.8 27B scores 33.7 on the same index, not far from OpenAI’s GPT-6 Luna at 38.1, which you rent.
The unlock speeds up prompt processing and image generation on cards you already own. It does not make text appear faster, and there is no evidence it helps games. If you try it, use a machine you can afford to break, read the project’s boot and recovery notes, and accept that you may lose your warranty and act against Nvidia’s licence terms. To undo it, remove the kernel option and power the machine off.
Honest limits
Krishna’s 70, 50 and 90 are not tied to a named benchmark, and the Artificial Analysis index is one yardstick among several. Prices are October 2026 list prices; real quotes include discounts. My cost arithmetic leaves out staff, cooling, networking and space. The GPU unlock results are community reports on one or a few cards per model.
Sources
- The Motley Fool, IBM Q3 2026 earnings call transcript (October 21, 2026)
- The Register, IBM CEO says open weight models are now irresistibly good and cheap on-prem (October 22, 2026)
- Artificial Analysis, open-weights model comparison and Intelligence Index v4.3.2 (read October 26, 2026)
- Artificial Analysis, Kimi K3.1 Pro launch benchmarks (October 23, 2026)
- DeepSeek, API pricing (read October 26, 2026)
- Anthropic, Claude API pricing (read October 26, 2026)
- Hugging Face, MiniMaxAI/MiniMax-M3.1-Flash (read October 26, 2026)
- BIZON, X9000 G3 8-GPU HGX H200 server (list price read October 26, 2026)
- NVIDIA, H200 specifications and DGX B300 specifications
- Tom’s Hardware, Nvidia raises RTX PRO 6000 Blackwell to $16,000 (August 12, 2026)
- RunsOn, AWS EC2 p5en prices from the AWS Pricing API (read October 26, 2026)
- US Energy Information Administration, Electric Power Monthly, Table 5.6.A (released October 23, 2026)
- Zagitta, gpu-unlock README and How Nvidia Ampere generation GPUs got unlocked
- JustAResearcher, gpu-unlock-ada README
- nlzy, turing-unlock README
- Hacker News, GPU-unlock discussion (October 24, 2026)
- Notebookcheck, RTX 3090 performance nearly doubles with free GPU unlock, but there’s a catch (October 24, 2026)
- MakeUseOf, a firmware unlock nearly doubles AI speed on RTX 30 and 40 GPUs (October 26, 2026)
- Wccftech, RTX 30 and RTX 40 GPUs get faster in AI with new unlock tool, but prices surge too (October 24, 2026)
- NVIDIA, GeForce software license (read October 26, 2026)
Researched and drafted with AI assistance, including Anthropic’s Claude.
