Why Salesforce Built Its Own Model, Koa
Large platforms are post-training open weights - often NVIDIA Nemotron - on their own synthetic and domain data, then hosting that model as a product brain they control. Salesforce Koa is the lead case. ServiceNow Apriel and Now LLM, Adobe's Firefly foundation, and the NVIDIA Nemotron substrate show the same move. Gartner's domain-specific forecast (1% of enterprise GenAI models in 2024 to more than half by 2027) is the scale proxy. There is no public Fortune-500 headcount. Do not invent one. The buyer question is build-own versus rent-frontier - not "fine-tune our CRM rows."
TLDR
What this research does. This page documents which platforms now own and operate their own foundation models for revenue work, and which still rent that layer from a frontier provider, with the sourcing for each.
Why it matters to you. Model ownership changes pricing, data flow, and roadmap control on the platforms you already run. This tells you which vendors can reprice or redirect your AI costs, and when.
The short answer
What does the research show about Why Salesforce Built Its Own Model, Koa?
Evidence
- Enterprise buyers mirror platforms: which workloads stay on Claude/OpenAI rent, which run on owned weights on Nvidia, which Order Form line pays for each path. Latham is law-domain own-model on firm-controlled Nvidia infra - not a CRM fine-tune story.
- Kill the myth: Koa is not "Salesforce trains on your CRM." Salesforce IR states no customer data was used to train Koa.
- Buyer question: which agent workflows stay on Claude/OpenAI rent, which move to a platform-owned model (Koa / Now LLM / peer), and which Order Form lines prove the bill on each path?
Supporting pages
- Dictionary plain-language definitions
Which big platforms now ship a brain they own, how that sits next to Claude and OpenAI rent, and the spend numbers that make the shift real. Every figure has a publisher and a date.
Claude and OpenAI still take most of the enterprise API dollar. Salesforce, ServiceNow, and Adobe are also building models they control so the product is not stuck on rented frontier as the only brain. Koa is Salesforce's version. Kill the myth once: Koa is not "Salesforce trains on your CRM." Then ask which of your agent workloads stay on rent and which should run on a platform-owned model.
Sep 15, 2026. Salesforce and NVIDIA announce Koa, Salesforce's first CRM reasoning model for Agentforce. Base: NVIDIA Nemotron 3 Super (open weights). Post-train: proprietary synthetic CRM corpus modeled on ~27 years of deployment knowledge (processes, workflows, policies) - not live customer org records. Ownership: Salesforce controls the weights; post-training and inference inside the Salesforce trust boundary. Status: select pilots now; GA expected Winter 2026, U.S. regions. Salesforce IR: "No customer data was used to train the Koa reasoning model." Vendor CRM Bench scores stay UNPROVEN here. Pilot vs GA and Flex Credit detail stays on the live decoder: Salesforce Koa pilot vs GA. Primary: Salesforce IR, Announcing Koa.
This page is about platforms building their own brains. Named vendor evidence. Spend and adoption proxies. Why buyers and platforms open to controlled or owned models. On-prem vs cloud hosting mix. Impact beyond unit price.
This page is not a guide to customer fine-tunes on org CRM rows. Not a RAG how-to. Not Missionforce customer-data post-training as the lead. Not Snowflake Cortex PEFT or M365 Copilot Tuning as the story. Those surfaces exist elsewhere; they are out of frame here.
Figure 1. Named platforms on the vendor-owned brain path
Public sources do not publish a Fortune-500 headcount. Use named evidence. Do not invent one.
Salesforce Koa
Synthetic CRM expertise corpus. Salesforce owns weights. IR: no customer data used to train Koa. Pilots now; GA Winter 2026 U.S.
Nemotron 3 Super
High confidence
ServiceNow Apriel + Now LLM
Post-trained with NVIDIA data and ServiceNow domain data. Proprietary Now LLM / Now SLM v2.0 for Now Assist. Customers can still route to Azure OpenAI, Gemini, Claude on Bedrock, or Now LLM.
Apriel Nemotron 15B
High confidence
Adobe Firefly foundation
Enterprise creative generative foundation Adobe owns and ships. Enterprise content stated as not used to train foundational Firefly. Customer brand Custom Models are a different product and out of frame here.
Adobe-owned foundation
Medium confidence
NVIDIA Nemotron
Koa and Apriel both sit on Nemotron open models. Platforms post-train with their synthetic or domain corpus, host as a managed LLM, and offer it beside Claude/OpenAI/Gemini.
Shared substrate
High confidence
What this does not say
This roster is not a complete market census. Absence from the list is not proof a vendor lacks any private model work.
- Publisher
- Salesforce IR, ServiceNow press and community, Adobe business pages
- Sample and method
- Named primary sources only. Workday Sana and HubSpot Breeze are not forced into the owns-a-foundation-brain column; no LOCKED primary in the research packet.
- Field dates
- ServiceNow Apriel May 6, 2025; Now LLM v2.0 rollout from Oct 16, 2025; Salesforce Koa Sep 15, 2026
Enterprise own-model peers on Nvidia
Same own-vs-rent move on the buyer side: firm- or enterprise-owned domain models on Nvidia infra, still hybrid with Claude/OpenAI/Gemini rent where needed. Not customer CRM fine-tunes. Latham & Watkins (~$8.3B U.S. #2 by revenue) spent roughly three years buying Nvidia GPU servers, including multiple H200s, and is fine-tuning open-weight Nvidia Nemotron 3 inside a data center only its staff can access. CIO Rene Mendoza (FT): keep sensitive client information off cloud vendors and keep flexibility as AI consumption pricing hardens. Michael Rubin (Bloomberg Law): on-prem open-weight path gives capabilities cloud legal AI tools do not. The firm still uses Harvey, ChatGPT, Claude, and Gemini. FT: first major law firm publicly known to stand up this class of in-house AI infrastructure (Sep 10-11, 2026). Do not invent Latham NIM/NeMo/DGX SKUs, spend, or GPU count. Peers below are named Nvidia case studies/blogs; ServiceNow Apriel is already in Figure 1.
Latham & Watkins
Law, Sep 2026. Nvidia H200 servers + open-weight Nemotron 3 on firm-controlled infra; hybrid with commercial LLMs. Sources: Bloomberg Law (lw.com PDF), FT via Legal IT Insider / Pulse2.
Nemotron 3 fine-tune
High confidence
RBC Capital Markets
Nemotron via NIM inside Aiden AI factory; hybrid on-prem + cloud. Nvidia case study.
Aiden + Nemotron via NIM
High confidence
Eli Lilly (LillyPod)
Wholly owned LillyPod DGX SuperPOD for protein/genomics/foundation models. Nvidia blog Feb 26, 2026.
Owned drug-discovery models
High confidence
Revolut
Own behavioral foundation model family on H100 clusters; Llama-Nemotron-Embed for text fields. Nvidia case study.
PRAGMA transaction FM
High confidence
Amgen
Antibody-specific protein LLMs on DGX Cloud + BioNeMo. Nvidia case study.
Proprietary protein LLMs
High confidence
AT&T
Ask AT&T agents; NeMo + NIM stack. Nvidia case study.
Fine-tuned agents
High confidence
Siemens
On-prem SLM inference on NIM / NeMo; shop-floor data stays on site. Nvidia case study.
On-prem Industrial Copilot
High confidence
What this does not say
Do not claim Latham uses NIM, NeMo, or DGX by SKU name. Do not claim Kirkland owns Nvidia servers. Do not invent spend, GPU count, or Fortune-500 headcounts.
- Publisher
- Bloomberg Law via lw.com; FT via Legal IT Insider / Pulse2; NVIDIA case studies and blogs
- Sample and method
- Buyer question stays the same as for Koa: which workloads stay on Claude/OpenAI rent, which move to a platform- or enterprise-owned model on Nvidia, and which Order Form line pays for each path.
- Field dates
- Latham Sep 10-11, 2026; peer Nvidia case studies as published
Figure 2. Frontier rent is still the default dollar
Enterprise LLM spend. Platforms building own models is a response to that concentration and that bill, not proof rent is gone. Menlo also: closed-source 87% of enterprise workloads mid-2025.
- Enterprise LLM spend ($B)4.9 pts
What this does not say
Do not mash Menlo's 13% open-source workload share with ETR's open-weight token share. Different metrics, different dates.
- Publisher
- Menlo Ventures Mid-Year LLM Market Update
- Sample and method
- n~150 technical leaders; fielded Jun 30-Jul 10, 2025; published Jul 31, 2025. Closed-source 87% of enterprise workloads; open-source 13%. Anthropic 32% / OpenAI 25% / Google 20% usage-weighted.
- Field dates
- Jun 30-Jul 10, 2025
Figure 4. Controlled open-weight tokens are already a real share
ETR via Techstrong. Read as proof buyers will run controlled models for volume work - the same economic logic platforms use when they post-train Nemotron instead of calling Claude/OpenAI for every step.
Share
- Open-weight share of enterprise AI tokens (now)34%
- Expected open-weight token share in 12 months41%
- Enterprises with open-weight in production42%
- Among those: self-managed cloud hosting46%
- Among those: on-prem hosting42%
- Production users citing cost savings69%
What this does not say
Platform-owned brains (Koa, Now LLM) are typically vendor-hosted managed LLMs, not customer-downloadable weights. ETR hosting mix is the enterprise side of controlled weights.
- Publisher
- ETR survey via Techstrong.ai
- Sample and method
- n=200; Sep 2026. Security/compliance review incomplete = 62% of pilot blockers.
- Field dates
- September 2026
ETR via Techstrong. Read as proof buyers will run controlled models for volume work - the same economic logic platforms use when they post-train Nemotron instead of calling Claude/OpenAI for every step.
| Share | Value (%) | Note |
|---|---|---|
| Open-weight share of enterprise AI tokens (now) | 34 | |
| Expected open-weight token share in 12 months | 41 | |
| Enterprises with open-weight in production | 42 | |
| Among those: self-managed cloud hosting | 46 | |
| Among those: on-prem hosting | 42 | |
| Production users citing cost savings | 69 |
Source: ETR survey via Techstrong.ai. n=200; Sep 2026. Security/compliance review incomplete = 62% of pilot blockers. Fielded September 2026. Confidence: High.
What this does not say: Platform-owned brains (Koa, Now LLM) are typically vendor-hosted managed LLMs, not customer-downloadable weights. ETR hosting mix is the enterprise side of controlled weights.
Figure 5. Token bills and platform scale both push off pure rent
EY puts token burn on the calendar. Salesforce ships massive frontier-style throughput and a parallel bet on Koa.
EY US AI Pulse (5th wave)
98% of senior leaders using token tools say token costs caused them to reconsider approach; 82% concerned. n=534 SVP+; fielded Apr 24-May 17, 2026; published Jul 28, 2026.
98% reconsidered
High confidence
Salesforce FY26 Q4
Agentforce ARR $800M (+169% Y/Y); >19T tokens all-time (5x Y/Y); 2.4B Agentic Work Units. Released Feb 25, 2026.
$800M Agentforce ARR
High confidence
What this does not say
Agentforce token volume does not prove Koa replaces frontier rent. Both can be true at once.
- Publisher
- EY; Salesforce IR
- Sample and method
- Two independent reads: buyer fiscal scrutiny and one platform's agentic scale while it builds its own brain.
- Field dates
- EY fielded Apr 24-May 17, 2026; Salesforce period ended Jan 31, 2026
Also in the record
Figures that sit alongside these charts.
- Enterprise buyers mirror platforms: which workloads stay on Claude/OpenAI rent, which run on owned weights on Nvidia, which Order Form line pays for each path. Latham is law-domain own-model on firm-controlled Nvidia infra - not a CRM fine-tune story.
- Kill the myth: Koa is not "Salesforce trains on your CRM." Salesforce IR states no customer data was used to train Koa.
- Buyer question: which agent workflows stay on Claude/OpenAI rent, which move to a platform-owned model (Koa / Now LLM / peer), and which Order Form lines prove the bill on each path?
- Platform-owned brains (Koa, Now LLM, Firefly foundation): vendor hosts. Weights stay with the platform. Enterprise-controlled open-weight (ETR hosting mix) is a different owner of the stack.
- Name the Order Form line before volume agents go live: rented tokens vs platform credits/AWUs vs fixed infra for self-hosted open-weight.
- Exit cost: swapping an API key is cheaper than leaving a platform-owned reasoning path baked into the product.
- Do not invent a Fortune-500 fine-tune headcount. Use named platform evidence plus Gartner/Menlo/ETR proxies.
Questions this page answers
What the data says, in plain language.
- What does the research show about Why Salesforce Built Its Own Model, Koa?
- Large platforms are post-training open weights - often NVIDIA Nemotron - on their own synthetic and domain data, then hosting that model as a product brain they control. Salesforce Koa is the lead case. ServiceNow Apriel and Now LLM, Adobe's Firefly foundation, and the NVIDIA Nemotron substrate show the same move. Gartner's domain-specific forecast (1% of enterprise GenAI models in 2024 to more than half by 2027) is the scale proxy. There is no public Fortune-500 headcount. Do not invent one. The buyer question is build-own versus rent-frontier - not "fine-tune our CRM rows." Which big platforms now ship a brain they own, how that sits next to Claude and OpenAI rent, and the spend numbers that make the shift real. Every figure has a publisher and a date.
- What does the figure "Figure 1. Named platforms on the vendor-owned brain path" show?
- Public sources do not publish a Fortune-500 headcount. Use named evidence. Do not invent one. Source: Salesforce IR, ServiceNow press and community, Adobe business pages. Named primary sources only. Workday Sana and HubSpot Breeze are not forced into the owns-a-foundation-brain column; no LOCKED primary in the research packet. Fielded ServiceNow Apriel May 6, 2025; Now LLM v2.0 rollout from Oct 16, 2025; Salesforce Koa Sep 15, 2026. Confidence: High.
- What does the figure "Enterprise own-model peers on Nvidia" show?
- Same own-vs-rent move on the buyer side: firm- or enterprise-owned domain models on Nvidia infra, still hybrid with Claude/OpenAI/Gemini rent where needed. Not customer CRM fine-tunes. Latham & Watkins (~$8.3B U.S. #2 by revenue) spent roughly three years buying Nvidia GPU servers, including multiple H200s, and is fine-tuning open-weight Nvidia Nemotron 3 inside a data center only its staff can access. CIO Rene Mendoza (FT): keep sensitive client information off cloud vendors and keep flexibility as AI consumption pricing hardens. Michael Rubin (Bloomberg Law): on-prem open-weight path gives capabilities cloud legal AI tools do not. The firm still uses Harvey, ChatGPT, Claude, and Gemini. FT: first major law firm publicly known to stand up this class of in-house AI infrastructure (Sep 10-11, 2026). Do not invent Latham NIM/NeMo/DGX SKUs, spend, or GPU count. Peers below are named Nvidia case studies/blogs; ServiceNow Apriel is already in Figure 1. Source: Bloomberg Law via lw.com; FT via Legal IT Insider / Pulse2; NVIDIA case studies and blogs. Buyer question stays the same as for Koa: which workloads stay on Claude/OpenAI rent, which move to a platform- or enterprise-owned model on Nvidia, and which Order Form line pays for each path. Fielded Latham Sep 10-11, 2026; peer Nvidia case studies as published. Confidence: High.
- What does the figure "Figure 2. Frontier rent is still the default dollar" show?
- Enterprise LLM spend. Platforms building own models is a response to that concentration and that bill, not proof rent is gone. Menlo also: closed-source 87% of enterprise workloads mid-2025. Source: Menlo Ventures Mid-Year LLM Market Update. n~150 technical leaders; fielded Jun 30-Jul 10, 2025; published Jul 31, 2025. Closed-source 87% of enterprise workloads; open-source 13%. Anthropic 32% / OpenAI 25% / Google 20% usage-weighted. Fielded Jun 30-Jul 10, 2025. Confidence: High.
- What does the figure "Figure 3. Domain-specific models are forecast to become the majority" show?
- Gartner: domain-specific share of enterprise GenAI models. Structural demand proxy for specialized or owned domain models. Not a count of CRM fine-tunes. Source: Gartner PR. GenAI model end-user spend $14.2B in 2025; specialized / domain models $1.1B in 2025. Domain-specific models: 1% of enterprise GenAI models in 2024 to more than half by 2027. Fielded Published Jul 10, 2025. Confidence: High. Caveat: Newsroom may bot-wall; keep primary URL. "More than half" charted at 51% for display only.
- What else sits alongside these figures?
- Enterprise buyers mirror platforms: which workloads stay on Claude/OpenAI rent, which run on owned weights on Nvidia, which Order Form line pays for each path. Latham is law-domain own-model on firm-controlled Nvidia infra - not a CRM fine-tune story. Kill the myth: Koa is not "Salesforce trains on your CRM." Salesforce IR states no customer data was used to train Koa. Buyer question: which agent workflows stay on Claude/OpenAI rent, which move to a platform-owned model (Koa / Now LLM / peer), and which Order Form lines prove the bill on each path? Platform-owned brains (Koa, Now LLM, Firefly foundation): vendor hosts. Weights stay with the platform. Enterprise-controlled open-weight (ETR hosting mix) is a different owner of the stack.
- Where does this data come from?
- Every figure is reproduced from a named publisher: Salesforce IR, Announcing Koa, Salesforce, Koa product page, The Revenue AI Report, Koa pilot vs GA decoder, ServiceNow x NVIDIA press, ServiceNow Community, AI Models v2.0, ServiceNow Community, AI SDLC Part 2. Sample, field date, and confidence are shown on each chart. Sources marked as vendor research are labelled on the page.
- What could not be confirmed?
- Salesforce CRM Bench claims (3x fewer errors; product-page lifts). Vendor benchmark only until third-party replication. Any public Fortune-500 headcount of companies that built a Koa-class vendor-owned foundation LLM. Do not invent one. Workday Sana or HubSpot Breeze shipping a Koa-class owned foundation LLM as the default brain. No LOCKED primary in the packet.
Cite this page
Permanent URL and suggested citation.
https://www.therevenueaireport.com/research/platform-owned-models-koa-nemotron-vs-claude-openai-rent
Kvarfordt, Jonathan. "Why Salesforce Built Its Own Model, Koa." The Revenue AI Report, Research Library. https://www.therevenueaireport.com/research/platform-owned-models-koa-nemotron-vs-claude-openai-rent
Figures on this page are reproduced from the publishers listed below. Cite the original publisher for the underlying data, and this page for the compilation and framing.
Sources
Every publisher used on this page.
If a metric, model term, or method on this page is unfamiliar, every one of them is defined in The AI and Revenue Dictionary. Sample size, field date, and confidence tags are explained there too.
Salesforce IR, Announcing Koa
First CRM reasoning model; Nemotron 3 Super; synthetic corpus; no customer data used to train Koa; Sep 15, 2026
https://investor.salesforce.com/news/news-details/2026/Announcing-Koa-Salesforces-First-CRM-Reasoning-Model-Built-on-NVIDIA-Nemotron/default.aspxSalesforce, Koa product page
Agentforce CRM reasoning model overview
https://www.salesforce.com/agentforce/koa/The Revenue AI Report, Koa pilot vs GA decoder
Pilot vs GA and Flex Credit detail; do not duplicate here
https://www.therevenueaireport.com/blog/salesforce-koa-agentforce-pilot-vs-gaServiceNow x NVIDIA press
Apriel Nemotron 15B; May 6, 2025
https://www.servicenow.com/company/media/press-room/nvidia-enterprise-ai-agents.htmlServiceNow Community, AI Models v2.0
Now LLM / Now SLM v2.0 rollout from Oct 16, 2025
https://www.servicenow.com/community/now-assist-articles/announcing-servicenow-ai-models-v2-0-with-enhanced-capabilities/ta-p/3405960ServiceNow Community, AI SDLC Part 2
CPT + SFT on proprietary / synthetic / vendor datasets
https://www.servicenow.com/community/now-assist-articles/part-2-inside-the-ai-sdlc-at-servicenow-from-foundational-models/ta-p/3304991Adobe Firefly Custom Models business page
Foundation / enterprise training claims; foundation is the peer; custom brand models out of frame
https://business.adobe.com/products/firefly-business/custom-models.htmlGartner newsroom, GenAI model spend forecast
Domain-specific 1% (2024) to more than half by 2027; $14.2B / $1.1B in 2025
https://www.gartner.com/en/newsroom/press-releases/2025-07-10-gartner-forecasts-worldwide-end-user-spending-on-generative-ai-models-to-total-us-dollars-14-billion-in-2025Menlo Ventures, 2025 Mid-Year LLM Market Update
$3.5B to $8.4B; closed 87% workloads; Jul 31, 2025
https://menlovc.com/perspective/2025-mid-year-llm-market-update/ETR via Techstrong.ai
Open-weight 34% to 41% of enterprise tokens; 42% in production; hosting 46%/42%; Sep 2026
https://techstrong.ai/articles/open-weight-ai-models-expected-to-capture-41-of-enterprise-token-usage/EY US AI Pulse
98% reconsidered / 82% concerned on token costs; Jul 28, 2026
https://www.ey.com/en_us/newsroom/2026/07/ey-survey-c-suites-pivot-from-ai-adoption-to-unlocking-value-as-escalating-token-costs-trigger-fiscal-scrutinySalesforce IR, FY26 Q4 results
Agentforce ARR $800M; >19T tokens; 2.4B AWUs; Feb 25, 2026
https://investor.salesforce.com/news/news-details/2026/Salesforce-Delivers-Record-Fourth-Quarter-Fiscal-2026-Results/Bloomberg Law via Latham (lw.com PDF)
Latham Nvidia GPU servers / H200; on-prem open-weight optionality; hybrid with commercial LLMs; Sep 11, 2026
https://www.lw.com/en/industries/admin/upload/SiteAttachments/Latham-Own-AI-Servers-Give-Firm-Best-Options-Leaders-Say.pdfLegal IT Insider (FT summary)
Latham builds own AI models with Nvidia GPU investment; Mendoza FT quotes; Sep 11, 2026
https://legaltechnology.com/latham-builds-its-own-ai-models-with-nvidia-gpu-server-investment/Pulse2
Latham Nemotron 3 fine-tune named; Sep 11, 2026
https://pulse2.com/latham-watkins-buys-nvidia-gpu-servers-to-build-in-house-ai-systems/NVIDIA, RBC Capital Markets case study
Aiden AI factory; Nemotron via NIM
https://www.nvidia.com/en-us/case-studies/rbc-capital-markets/NVIDIA blog, Lilly AI factory
LillyPod DGX SuperPOD; Feb 26, 2026
https://blogs.nvidia.com/blog/lilly-ai-factory-live/NVIDIA, Revolut case study
PRAGMA on H100; Llama-Nemotron-Embed
https://www.nvidia.com/en-us/case-studies/revolut/NVIDIA, Amgen case study
Proprietary protein LLMs on DGX Cloud + BioNeMo
https://www.nvidia.com/en-gb/case-studies/amgen-biologics-discovery-and-development/NVIDIA, AT&T case study
NeMo / NIM agents
https://www.nvidia.com/en-gb/case-studies/att-drives-ai-agents-with-nemo/NVIDIA, Siemens case study
On-prem Industrial Copilot; NIM / NeMo
https://www.nvidia.com/en-us/case-studies/siemens-accelerates-product-development-and-innovation-with-industrial-ai/The Revenue AI Report subscribe
Soft CTA only
https://www.therevenueaireport.com/subscribe
Could not confirm
What we looked for and did not find.
Claims found during research and not charted
- Salesforce CRM Bench claims (3x fewer errors; product-page lifts). Vendor benchmark only until third-party replication.
- Any public Fortune-500 headcount of companies that built a Koa-class vendor-owned foundation LLM. Do not invent one.
- Workday Sana or HubSpot Breeze shipping a Koa-class owned foundation LLM as the default brain. No LOCKED primary in the packet.
How to cite this research
Written by Jonathan Kvarfordt, Founder and Principal Analyst, The Revenue AI Report. Published under CC BY 4.0.
APA
Kvarfordt, J. (2026). Why Salesforce Built Its Own Model, Koa. The Revenue AI Report. Retrieved from https://www.therevenueaireport.com/research/platform-owned-models-koa-nemotron-vs-claude-openai-rent
MLA
Kvarfordt, Jonathan. "Why Salesforce Built Its Own Model, Koa." The Revenue AI Report, 31 Aug. 2026, www.therevenueaireport.com/research/platform-owned-models-koa-nemotron-vs-claude-openai-rent.
BibTeX
@misc{kvarfordt2026platformownedmodelskoanemotronvsclaudeopenairent,
author = {Kvarfordt, Jonathan},
title = {Why Salesforce Built Its Own Model, Koa},
year = {2026},
publisher = {The Revenue AI Report},
url = {https://www.therevenueaireport.com/research/platform-owned-models-koa-nemotron-vs-claude-openai-rent}
}Next theme
AI ROI Starts With the OutcomeThe evidence points to one sequence: choose the strategic outcome, decompose the metric, inspect the operating system beneath it, test whether AI fits, then choose the OAR ambition and measure the result.
Subscribe