Tuesday, November 17, 2026 AI news, turned into opportunities PDF newsletter
AI Opportunity Daily
Subscribe
Open modelsData centresFundingEnergy

Kimi's 2.8T open model, Google's $17bn Wyoming bet, Thinking Machines eyes $40bn

Moonshot made one of the largest AI models ever released downloadable, Google committed $17 billion to a Wyoming data centre that brings its own gas-and-carbon-capture power, and Mira Murati's Thinking Machines is reported to be close to a $5 billion round. The thread: strong open models are getting cheaper to obtain while the power, money and skills needed to run them are getting more concentrated.

Key takeaways

  • Kimi K3.1 Pro's open weights put a model that Moonshot says is close to Claude Opus 5.5 on several coding tests within reach of any business that can host it, under a mostly permissive licence.
  • Google's $17 billion Wyoming campus shows big AI buyers now build their own power, here a 702 MW gas fuel-cell plant with carbon capture, rather than wait for the grid.
  • Investors still back AI labs at huge multiples: Thinking Machines is reported to be raising about $5 billion at a $40 billion pre-money valuation, a day after shipping Inkling 2 Small.
Story 1 of 3Models5 min read

Moonshot puts its 2.8-trillion-parameter Kimi K3.1 Pro on Hugging Face as open weights

Moonshot AI published the full weights of Kimi K3.1 Pro on November 4, a 2.8-trillion-parameter mixture-of-experts model with a one-million-token context window that the company says rivals Claude Opus 5.5 on several coding benchmarks.

Moonshot AI, the Beijing-based maker of the Kimi chatbot, published the full weights of Kimi K3.1 Pro on Hugging Face on November 4, keeping a promise it made when it launched the model through its API on October 23. The repository, moonshotai/Kimi-K3.1-Pro, is public and does not require an access form, according to coverage of the release. The model card calls it the company's most capable model to date. That makes it one of the largest models anyone can download, inspect, fine-tune and run on their own hardware, rather than only renting it through someone else's API.

The numbers are large. Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters, of which about 104 billion are active for each token, spread across 896 experts with 16 chosen per token. It reads up to about one million tokens of context and accepts images through a 401-million-parameter vision encoder. The weights ship in a compressed MXFP4 format, and the files total roughly 1.56 terabytes, according to NeuralBrew. Moonshot also sells access through its own API; Simon Willison reported standard pricing of $2 per million input tokens and $10 per million output tokens.

On Moonshot's own model card, K3.1 Pro scores 65.7 on Terminal-Bench 4.0 against 66.4 for Anthropic's Claude Opus 5.5, and 75.5 on DeepSWE against 74.2 for Opus 5.5 and 74.1 for OpenAI's GPT-6 Astra. It also posts 92.1 on the BrowseComp web-research test. But the same card shows clear gaps: 31.5 on Program Bench against 65.0 for Opus 5.5, and 39.4 on HLE-Diamond against 60.9. The picture is a model that sits close to the frontier on many coding and agent tasks, and well behind it on others.

All of these scores are reported by Moonshot, and the card notes that many comparison numbers come from other vendors' reports or leaderboards, with some tests run internally. Independent checks are still coming in. The licence is a custom Kimi K3.1 Pro License. NeuralBrew describes it as largely MIT-style with two conditions: companies reselling API access with more than $20 million in revenue over 12 months need a separate agreement, and products above 100 million monthly users or $20 million in monthly revenue must display the model's name. NeuralBrew also reports that Moonshot removed cybersecurity data from the open version's training; we could not confirm that on the model card.

For most small teams, the practical point is not running 1.56 terabytes of weights at home, which needs data-centre-class hardware. It is that a near-frontier model can now be offered by many hosting companies and deployed privately by any firm with the right servers or cloud contract. That puts pressure on closed-model prices, gives regulated businesses an option that keeps data in-house, and turns 'which model, on which host, at what cost' into a real question that clients will pay someone to answer. Expect inference providers to compete on price for it in the coming weeks.

Why it mattersOne of the strongest models anyone can download now comes under a mostly permissive licence, giving businesses a serious alternative to renting closed models for coding, research and agent work.

👀 What to watch

Independent benchmark results, which hosting providers offer the model and at what price, and whether large resellers run into the licence conditions.

3 opportunities from this story

1

Full-tender review service for small contractors

Low⏱ 1-2 weeks💰 Fixed fee per tender reviewed, with a higher tier that includes drafting the compliance matrix

Government and corporate tenders often run to hundreds of pages, and small contractors miss mandatory documents or scoring criteria buried in annexes. A model with a one-million-token context can read a whole tender pack in one go and produce a checklist of requirements, deadlines and evaluation criteria. Offer a fixed-price review per tender that a person checks before delivery. Kimi K3.1 Pro's API pricing makes each review cheap to run, and you can switch to another long-context model if quality or price changes.

Best for
Freelance bid writers, procurement consultants and virtual assistants who serve construction, IT and supply firms
First step this week
Download one public tender this week, run the full pack through a long-context model, and compare its checklist against your own manual read.
Open the full playbook
Launch steps
  1. Collect three past public tenders in your target sector
  2. Write one reusable prompt that extracts mandatory documents, deadlines, eligibility rules and scoring criteria
  3. Check every output by hand and note what the model missed
  4. Turn the checked output into a one-page sample report
  5. Offer the first two reviews at a discount to contractors you already know
Tools
Kimi K3.1 Pro API or another long-context modelA PDF-to-text converterGoogle Sheets for the compliance checklist
Risks

Models can miss or invent requirements, and a missed document can disqualify a bid, so a human must check every checklist and your terms should limit liability. Tender documents can be confidential, so confirm whether the client allows sending them to an outside API, and use a provider whose data terms you have read.

2

Private open-model setup for data-sensitive firms

High⏱ 4-8 weeks💰 Setup fee plus monthly management fee, with hosting costs passed through or marked up

Law firms, clinics, banks and NGOs often want strong AI but cannot send client data to a public chatbot. With near-frontier open weights available, you can arrange a dedicated deployment through a cloud or inference provider in an approved region, wire it to a simple internal chat or document tool, and document where the data goes. Sell the setup plus a monthly support plan. The value is the paperwork and plumbing, not the model itself.

Best for
DevOps and ML consultants, managed IT providers and cloud resellers
First step this week
Price a dedicated endpoint for an open model with two inference providers or cloud platforms, and write a one-page data-flow diagram you could show a client.
Open the full playbook
Launch steps
  1. Pick one sector and list its data rules (for Kenya, start with the Data Protection Act 2019)
  2. Compare dedicated-hosting options and regions for a large open model, plus a smaller fallback model
  3. Build a pilot with an internal chat front end, access logs and no training on client data
  4. Write a plain-English data-handling note the client's compliance officer can sign off
  5. Run a paid four-week pilot with one client before offering a long contract
Tools
vLLM or SGLangA cloud GPU or dedicated inference providerOpen WebUIAn access-logging tool
Risks

A 2.8-trillion-parameter model is expensive to host, so many clients will be better served by a smaller open model; quote honestly. Licence terms, export rules and client security reviews can slow deals. You become responsible for uptime and security, so carry professional liability cover and keep a fallback provider.

3

AI bill audit for start-ups

Medium⏱ 2-3 weeks💰 Fixed-fee audit, optionally with a share of first-year savings or a follow-on implementation contract

Many start-ups send every request to the most expensive closed model by default. A short audit can show which tasks could move to cheaper open models such as Kimi K3.1 Pro, or to smaller models, without hurting quality. You test their real prompts on two or three alternatives, measure quality and cost, and hand over a routing plan. Clients understand a report that says 'this change could save this much a month'.

Best for
Freelance developers and AI engineers who already build with LLM APIs
First step this week
Take 30 real prompts from one of your own projects, run them on your current model and one open model, and record cost and quality side by side.
Open the full playbook
Launch steps
  1. Ask the client for a month of API usage and a sample of anonymised prompts
  2. Group requests by task type and current cost
  3. Test each group on two or three alternative models with a simple scoring sheet
  4. Recommend which tasks to move, and estimate monthly savings with ranges
  5. Offer to implement the routing and monitor quality for a month
Tools
OpenRouter or a similar model routerPromptfooA spreadsheet for cost and quality scores
Risks

Savings can vanish if quality drops and users complain, so always present ranges and test on real traffic first. Prices change often, so date your report. Some clients will be wary of Chinese-developed models for data or policy reasons; respect that and offer alternatives.

Story 2 of 3Hardware5 min read

Google commits $17bn to its first Wyoming data centre, powered by a 702 MW gas fuel-cell plant with carbon capture

Google announced on November 5 a $17 billion data centre in Laramie County, Wyoming, its first in the state, with power partly supplied by a 702-megawatt natural-gas fuel-cell plant that captures and stores most of its carbon emissions.

Google announced on November 5 that it is building its first data centre in Wyoming, a $17 billion investment in Laramie County near Cheyenne. In a company blog post, Steve Walter, Google's vice-president for global infrastructure and energy, named four partners: Tallgrass Energy, Black Hills Energy, American Electric Power (AEP) and Related Digital. The headline feature is the power supply. Instead of relying only on the existing grid, the campus will draw on a large dedicated fuel-cell plant and gas generators, with much of the carbon dioxide captured and stored underground rather than released into the air.

Google says it has contracted with AEP for a 702-megawatt fuel-cell development with carbon capture and storage, which it calls the largest planned project of its kind to date. The fuel cells will be made in the United States by Bloom Energy and run on natural gas. Captured carbon will be stored at Tallgrass's Eastern Wyoming Sequestration Hub. Google adds that the deal brings its total contracted carbon-capture power to nearly 1 gigawatt. Carbon Herald reports that commercial operation of the first fuel-cell phase is targeted for late 2028, a date Google's own post does not give.

The site will be air-cooled to limit water use, which Carbon Herald notes uses roughly 10% more power than wet cooling. Google says it will pay for 100% of the power it uses and cover infrastructure needs driven by its growth, and that the project will bring hundreds of millions of dollars in property taxes to Laramie County and its school district. The timing stands out: the announcement came two days after US local elections in which, as we reported on November 3, data centres were on 41 ballots, and amid wider investor worry about local resistance to new campuses.

Several things remain unclear. Google says the system will capture 'a large majority' of emissions but gives no percentage, and Interesting Engineering stresses that the plant still runs on natural gas and reduces emissions rather than eliminating them. Carbon capture attached to a data centre at this scale has not yet been proven commercially, and costs, capture rates and timelines have not been fully published. Google's post gives no job count and no dollar figure for its community programmes. The project also depends on permits, gas supply and the sequestration hub working on schedule.

The bigger signal is that the largest AI buyers are now building their own power rather than waiting for grids to catch up. Gas plus carbon capture, sited next to the servers, is becoming a template, which matters for any region, including East Africa, that is courting data centres with promises of cheap or green electricity. For suppliers, the spending is shifting to everything around the chips: power equipment, monitoring, cooling, construction and skilled technicians. For communities, the open questions are who pays for grid upgrades and what happens to local water and air.

Why it mattersAI's growth is now limited by electricity, and Google's answer is to build dedicated gas power with carbon capture next to its servers, a model other hyperscalers and host countries will study closely.

👀 What to watch

The capture rate Google and AEP actually commit to, permit and pipeline milestones, and whether other hyperscalers copy the on-site fuel-cell approach.

3 opportunities from this story

1

Short courses for data-centre technicians

Medium⏱ 6-10 weeks💰 Per-student course fees, employer-sponsored cohorts and placement fees

Every new AI campus needs electricians, cooling technicians and power-plant operators, and the shortage of these people is a common bottleneck. Training colleges and experienced engineers can package short, practical courses on data-centre electrical systems, cooling, generators and safety, aimed at people who already hold a trade certificate. In Kenya and the wider region, where several data-centre projects are planned, employers and contractors need trained staff locally. Partnering with an operator for placements makes the course far easier to sell.

Best for
TVET colleges, experienced electrical and HVAC engineers, and training companies
First step this week
Call two local data-centre operators or contractors this week and ask which three skills they struggle to hire for.
Open the full playbook
Launch steps
  1. Interview operators and contractors about skill gaps and entry requirements
  2. Design a four-to-six-week course around those gaps, with hands-on modules
  3. Secure a site visit or placement partner
  4. Seek accreditation or an industry endorsement where possible
  5. Run a small paid pilot cohort and track job placements
Tools
Uptime Institute and vendor training material as referencesA learning platform such as MoodleLab equipment or partner site access
Risks

Courses without hands-on access or employer links struggle to place graduates, so line up partners first. Equipment for practical training is expensive. Data-centre build-outs can be delayed or cancelled, so make the skills transferable to other industrial sites.

2

Data-centre impact briefings for councils and landowners

Medium⏱ 3-4 weeks💰 Fixed-fee briefings, workshop fees, and retainers for ongoing project monitoring

Councils, community groups and landowners facing a data-centre proposal often lack independent information on power, water, tax and jobs. A consultant or researcher can produce a plain-English briefing that sets out what similar projects promised, what they delivered, and the questions to ask the developer. Google's Wyoming post, with its claims on power payments, air cooling and property taxes, is a useful template of what to ask for in writing. Buyers include local governments, community associations and the developers themselves, who want smoother approvals.

Best for
Policy researchers, planning consultants and energy journalists
First step this week
Build a one-page comparison of five recent data-centre announcements: investment, power source, cooling, stated tax and job benefits.
Open the full playbook
Launch steps
  1. Collect public filings and announcements for recent data-centre projects
  2. Create a standard checklist of questions on power, water, grid costs, tax and jobs
  3. Write a sample briefing on one real or hypothetical proposal
  4. Pitch it to a local government or community group near a planned site
  5. Offer a follow-up workshop for officials or residents
Tools
Public planning and utility filingsGoogle SheetsCanva for one-page summaries
Risks

You must stay independent and transparent about who pays you, or your credibility goes. Public data may be incomplete, so state what you could not verify. Demand depends on projects in your area.

3

AI energy and carbon reporting for mid-sized firms

Medium⏱ 2-4 weeks💰 Fixed fee per report, with an annual retainer for updates

As AI use grows, auditors, investors and big customers are starting to ask suppliers about the energy and emissions of their cloud and AI use. Most mid-sized firms have no idea how to answer. You can pull figures from cloud providers' carbon dashboards, estimate AI workloads with clear assumptions, and produce a short annual or quarterly report. Stories like Google's carbon-capture plant make the topic visible, and buyers want a defensible number rather than a guess.

Best for
Sustainability consultants, accountants and data analysts
First step this week
Open the carbon dashboards of the main cloud providers and write a one-page method for estimating a client's AI-related emissions.
Open the full playbook
Launch steps
  1. Learn the main reporting frameworks your clients' customers ask about
  2. Map where a client's AI and cloud usage data lives
  3. Pull provider carbon data and fill gaps with stated estimates
  4. Write a short report with methods, numbers and reduction options
  5. Offer quarterly updates as usage changes
Tools
Google Cloud Carbon FootprintMicrosoft Emissions Impact DashboardAWS Customer Carbon Footprint ToolCloud Carbon Footprint (open source)
Risks

Emissions estimates for AI APIs are uncertain and providers disclose different things, so be explicit about assumptions. Rules and frameworks change, so keep up to date. Some clients only want a box ticked; decline work that would require overstating accuracy.

Story 3 of 3Business5 min read

Mira Murati's Thinking Machines nears a $5bn round at a $40bn pre-money valuation, a day after shipping Inkling 2 Small

Thinking Machines Lab is close to raising about $5 billion led by Accel at a $40 billion pre-money valuation, Business Insider reported on November 6, the day after the start-up released its Inkling 2 Small model.

Thinking Machines Lab, the San Francisco start-up led by former OpenAI chief technology officer Mira Murati, is close to raising about $5 billion in a round led by Accel at a $40 billion pre-money valuation, Business Insider reported on November 6, according to Techstrong.ai and Tech Funding News. The round could close as soon as this week. If it does, the company would be valued at about $45 billion once the new money is counted. Neither side has confirmed the deal: Dataconomy says a company representative declined to comment and Accel did not respond.

The new price is more than three times the company's last mark. Thinking Machines raised a $2 billion seed round in July 2025 at a $12 billion valuation, from investors including Andreessen Horowitz, Accel, Nvidia, AMD, ServiceNow, Cisco and Jane Street. It is also below an earlier ambition: Dataconomy and Tech Funding News report that the company sought a valuation of around $50 billion last year before those talks were abandoned. Nvidia was separately reported to be in talks to invest about $2.5 billion, and in March the two companies agreed to deploy at least one gigawatt of Nvidia Vera Rubin systems starting in 2027.

The funding report landed a day after a product launch. On November 5 the company introduced Inkling 2 Small, a mixture-of-experts model with 594 billion total parameters and 17 billion active, which accepts images and lets users choose how much reasoning effort it applies. It is available through Tinker, the company's fine-tuning and inference service, and an online playground. Techstrong reports the weights are closed for now, with a release under the Apache 2.0 licence planned for next week after external safety testing. On the Artificial Analysis Intelligence Index it scores 31, up 18 points on its predecessor.

This is a report about a round that has not closed, and the terms could still change. Thinking Machines has disclosed little about its finances. A source told TechCrunch in September that annual revenue exceeded $100 million, which Tech Funding News works out to roughly 400 times revenue at the new valuation. The company has also lost people: Dataconomy reports about a third of its founding team has left, including co-founders who went to Meta and OpenAI. And Inkling 2 Small is not cheap to run: Techstrong says it uses about four times the output tokens per task, averaging $1.18 per index task against $0.20 for the earlier Inkling Small.

Even so, investors are still writing very large cheques for labs that sell tools to build on, not only finished chatbots. Thinking Machines' main business is Tinker, which charges customers to fine-tune models, and its strategy is to release capable open-weight models that those customers can customise. With a 594-billion-parameter model heading for an Apache 2.0 licence, and Moonshot's 2.8-trillion-parameter Kimi K3.1 Pro already downloadable, the supply of strong open models is growing quickly. The scarcer skill is turning them into specific, reliable tools for one business, which is where smaller builders can compete.

Why it mattersA $40 billion pre-money price for a lab with reported revenue of about $100 million shows investors are betting that open, customisable models plus fine-tuning tools will be a large business.

👀 What to watch

Whether the round closes on the reported terms, the size of any Nvidia investment, and whether Inkling 2 Small's weights ship under Apache 2.0 next week as planned.

3 opportunities from this story

1

Niche fine-tuning service for one business task

Medium⏱ 3-6 weeks💰 Project fee per fine-tuned model plus a monthly fee for hosting, monitoring and retraining

Fine-tuning services like Tinker make it practical to adapt open models to a narrow task, such as classifying customer messages, extracting fields from invoices or answering in a company's tone. Many small and mid-sized firms have the data but not the skills. Pick one task in one industry, build a repeatable fine-tuning and evaluation process, and sell it as a packaged result. A small fine-tuned model can be cheaper and more consistent than prompting a large general model for every request.

Best for
ML engineers, data scientists and technical freelancers
First step this week
Choose one task you know well, gather 200 example inputs and correct outputs, and test a fine-tuned small open model against a prompted large model.
Open the full playbook
Launch steps
  1. Pick one industry and one high-volume, repetitive task
  2. Collect and clean labelled examples with the client's permission
  3. Fine-tune a small open model and compare it with a prompted baseline on a held-out test set
  4. Deploy behind a simple API with logging
  5. Retrain monthly as new labelled data arrives
Tools
Tinker or another fine-tuning serviceHugging FaceUnslothLabel Studio
Risks

Fine-tuning only pays off with enough good data; many clients have messy or too little data, so check before quoting. Base models improve fast and may make your fine-tune unnecessary, so sell outcomes and keep switching costs low. Handle client data under a written agreement.

2

African-language evaluations of new open models

Medium⏱ 3-5 weeks💰 Paid detailed reports, custom evaluations for companies, and newsletter sponsorships

New open models arrive every week, but few are tested on Swahili, Sheng, Amharic, Yoruba or other African languages. Companies building local chatbots and support tools need to know which model handles their language best. Build a small, high-quality test set for one or two languages, run each major open release through it, and publish free headline results with paid detailed reports. Inkling 2 Small's planned Apache 2.0 release and Kimi K3.1 Pro give you two fresh models to test this month.

Best for
Linguists, NLP researchers and developers who are native speakers
First step this week
Write 100 realistic test questions and answers in one language, covering customer service, local facts and translation.
Open the full playbook
Launch steps
  1. Design a test set with native speakers and clear scoring rules
  2. Run three to five open models through it via hosted APIs
  3. Publish a free summary table and method note
  4. Sell detailed per-task reports and custom tests on client data
  5. Update with each major model release
Tools
Hugging FaceOpenRouterPromptfoo or lm-evaluation-harnessGoogle Sheets
Risks

Small test sets can mislead, so state their limits and grow them over time. Results go stale quickly. Model makers may contest findings, so publish your method and keep your scoring consistent.

3

AI start-up due-diligence notes for angel investors

Medium⏱ 2-3 weeks💰 Fixed fee per technical review, paid by the angel network, fund or lead investor

Angel networks and small funds are being pitched AI start-ups at high valuations, and many members cannot judge the technical claims. A short due-diligence note can check what the product depends on, whether its models or data are defensible, what it costs to run and how exposed it is to model price changes. Rounds like Thinking Machines' make valuations a talking point, and smaller investors want a second opinion before writing cheques. This is analysis, not investment advice.

Best for
Experienced engineers, technical product managers and AI consultants
First step this week
Write a two-page sample note on a public AI start-up using only its website, docs and pricing page.
Open the full playbook
Launch steps
  1. Create a standard checklist: model dependency, data rights, unit costs, security and team
  2. Write a sample note on a public company to show your format
  3. Approach local angel networks and accelerators
  4. Agree a fixed scope and confidentiality terms before each review
  5. Keep a record of past calls to show your track record
Tools
A due-diligence checklist templateNotion or Google DocsPublic pricing pages and API docs
Risks

You are giving technical opinions, not financial advice; say so clearly in your terms. Founders may share confidential information, so sign NDAs. A wrong call can hurt your reputation, so stay within what you can verify and list your uncertainties.

Get these as a PDF every 3 days

Free. One email every 3 days. Unsubscribe any time.

More editions