OpenAI trims its round and merges your limits, as Aramco buys half of HUMAIN
OpenAI spent the weekend getting smaller in two directions at once: it cut its fundraising target by $10bn and told paying users that from December their ChatGPT and Codex usage will come out of one shared token allowance. In the Gulf, money moved the other way, with Aramco closing a $2.6bn purchase of 49% of the Saudi AI company HUMAIN. The thread running through all three: the cheap, abundant phase of AI is being replaced by budgets, meters and state balance sheets.
Key takeaways
- From December, Plus, Pro and Business subscribers get one token allowance covering both ChatGPT and Codex, and OpenAI has not published how many tokens that is.
- OpenAI cut its round from $30bn to $20bn and its valuation from $1.4tn to $1.25tn, with about $11bn committed so far according to the Wall Street Journal.
- Aramco now owns 49% of HUMAIN, which plans 3 gigawatts of AI data-centre capacity by 2030, putting Saudi state capital behind compute while private AI funding tightens.
OpenAI will meter ChatGPT and Codex from one shared token allowance starting in December
OpenAI's help page now says that from December, Plus, Pro and Business plans will have a single usage allowance shared between ChatGPT and Codex, measured in tokens rather than messages. The company has not published how large that allowance is for any plan.
OpenAI updated its "Understanding ChatGPT and Codex usage limits" help article over the weekend with a short notice about a change due next month. In its own words: "Starting in December, Plus, Pro, and Business plans' usage allowance will be shared across ChatGPT and Codex as we bring the agentic capabilities of Work into the main ChatGPT experience." Until now, chat use and coding-agent use sat in different buckets for most subscribers. From December they come out of the same pot, and the page frames this as a side-effect of folding the Work product's agent features back into ordinary ChatGPT.
The second change is the unit of account. The page says usage "will be measured in tokens, the units of information a model processes and generates", and that "more advanced models and complex tasks generally use more of your allowance". Tokens are roughly word-fragments, so a long document pasted into a prompt now costs the same kind of resource as an hour of agent work on a repository. One exclusion is named: "Instant chat will not draw from your usage." OpenAI also promises visibility, saying users "will be able to see your combined remaining usage and reset times in Settings → Usage".
What OpenAI does not say is the part that matters most for anyone planning a month of work. There is no exact start date beyond "December", no table of how many tokens each plan gets, and no conversion showing how today's message-based limits translate into the new token figures. The page offers a reassurance rather than a number: OpenAI says it expects "most users won't need to wait for their usage to reset more often than they do now". If a subscriber does hit the ceiling, the options listed are to wait for the reset, upgrade the plan, or buy credits to continue using eligible features.
Coverage over the weekend agreed on the substance and differed on the framing. Crypto Briefing, writing on 31 October, described it as ChatGPT and Codex moving to a single token allowance in December. A detailed write-up by Pasquale Pillitteri on 1 November, which says the help page was updated on 31 October, argues the shift is narrower than it looks: Codex and the Work product already drew on a common pool, so the real change is that everyday ChatGPT chat now enters the same accounting. Both agree the plans named are Plus, Pro and Business. Neither found any mention of Free, Go, Enterprise or Edu plans, and OpenAI has issued no blog post or statement beyond the help page.
For anyone who bills for AI-assisted work, the practical consequence is a budgeting problem rather than a price rise. A $20 Plus subscription that comfortably covered a few hours of chat plus some agent runs may now behave differently once both draw on one meter, and heavy coding sessions will visibly eat into the budget for research and drafting. The change also arrives alongside OpenAI's funding squeeze, which makes a tightening of per-seat economics easier to read. Until the token figures are published, the safe assumption is that nothing is known about the new ceiling, and the sensible response is to measure your own consumption before December rather than after.
Why it mattersAnyone who resells AI-assisted work has been pricing jobs against a message allowance that is about to be replaced by a token budget they cannot yet size. That turns capacity planning from a known constraint into an unknown one for a month.
👀 What to watch
Whether OpenAI publishes per-plan token figures and a conversion from today's limits before December, and whether Free, Go, Enterprise and Edu plans are brought into the same scheme.
3 opportunities from this story
One-week AI token audit for small agencies
Sell a short, fixed-price audit that measures where a team's AI spend actually goes: which prompts, which models, which tasks. The deliverable is a one-page breakdown of consumption per workflow plus three concrete changes that cut usage without cutting quality, such as moving long-document summarising to a cheaper model. Your buyers are design, content, dev and marketing shops with three to twenty people who pay for several seats and have never looked at the numbers. The December change gives you a reason to call this week rather than next quarter.
- Best for
- Freelance operations consultants, technical VAs, agency ops leads
- First step this week
- Run the audit on your own workflow first: log every AI task for five days with the model used and rough prompt size, then write it up as a sample report.
Open the full playbook
Launch steps
- Build a simple logging sheet: date, task, model, approximate input and output size, who ran it
- Define three standard swaps you recommend (smaller model for drafts, batch instead of ad-hoc, reusable prompt templates)
- Produce a one-page sample report from your own data
- Offer the audit to five agencies you already know at an introductory price
- Add a follow-up review 30 days after the December change lands
Tools
Risks
Token allowances are still unpublished, so avoid promising savings in percentages. Sell measurement and a plan, not a guaranteed number, and say openly that the recommendations may need revisiting once OpenAI publishes figures.
Model-routing setup so cheap work stops hitting the premium meter
Many teams send every task to their most capable model out of habit. Build them a simple routing layer: classification, extraction, tagging and first drafts go to a small or open model, while the premium model handles only reasoning-heavy work. This is a two-to-four-week technical engagement with a measurable before-and-after. It sells best to teams running AI inside a product or an internal tool rather than just chatting in a browser.
- Best for
- Backend and automation developers, small dev shops
- First step this week
- Take one real workflow you already run, split it into 'needs reasoning' and 'does not', and benchmark the cheap model on the second half.
Open the full playbook
Launch steps
- Inventory the client's AI calls and group them by task type
- Benchmark two cheaper options per group against the current output on 20 real examples
- Implement routing behind a single internal interface so the client can change models later
- Add logging so the saving and the quality are both visible
- Hand over a short runbook naming which tasks go where and why
Tools
Risks
Cheaper models fail in ways that are hard to spot, so budget real time for evaluation on the client's own examples. Keep a one-line fallback to the premium model and be clear that some tasks will stay expensive.
Prompt and context hygiene training for non-technical teams
When usage is counted in tokens, habits start to cost money: pasting whole PDFs, re-sending context in every message, asking a reasoning model to reformat a list. Run a short practical workshop that teaches a sales, HR or finance team to get the same output with less input, and leave them with templates for their five most common tasks. This is a repeatable half-day product you can deliver remotely and sell to the same client every time they hire.
- Best for
- Trainers, corporate facilitators, consultants already inside companies
- First step this week
- Collect the ten worst prompts from your own clients or colleagues and rewrite each one into a reusable template with the context trimmed.
Open the full playbook
Launch steps
- Pick one department and document its five most frequent AI tasks
- Build a before-and-after example for each, showing shorter input and equal or better output
- Package a 90-minute session plus a template document
- Pilot it free with one friendly team in exchange for a testimonial
- Sell it to three similar departments or companies
Tools
Risks
Training is easy to sell once and hard to sell twice, so attach a follow-up review or a template subscription. Avoid teaching tricks tied to one provider's current interface, since the December change will move things around.
OpenAI cuts its funding round to $20bn and its valuation to $1.25tn
The Financial Times reported on Friday 30 October that OpenAI has cut its current fundraising target from $30bn to as much as $20bn, with the valuation falling from $1.4tn to $1.25tn. The Wall Street Journal put the amount committed so far at about $11bn.
OpenAI has reduced the size of the round it is currently raising. According to the Financial Times, first reported on Friday 30 October and picked up over the weekend, the target has fallen by $10bn, from $30bn to as much as $20bn. The valuation attached to the round has dropped with it, from $1.4tn to $1.25tn, a reduction of about $150bn. Friday was also the deadline for investors to make binding commitments. CNBC confirmed the $20bn figure on Saturday 31 October, citing a person directly familiar with the deal who stressed that the round is still open and the numbers could change again.
The Wall Street Journal reported that OpenAI has secured about $11bn so far, less than half the $30bn it originally sought, and described interest from some large investors as tepid. Reporting on who is in points in one direction: investors from the United Arab Emirates, including the state-backed fund MGX, are expected to be the largest participants, with SoftBank, Founders Fund, Coatue, Sequoia Capital and Andreessen Horowitz also named as planning to take part. Trending Topics adds that wealthy individuals would then be offered the same valuation, and that the first close could come in below the revised $20bn target.
The timing is not accidental. Anthropic's public listing is expected in mid-November at a valuation of at least $1.5tn, and its IPO filing became public last week. Both companies are courting the same institutional investors, so OpenAI has been trying to lock in commitments before Anthropic reaches the market. Reporting on OpenAI's own listing plans is less settled: the Journal says OpenAI plans to list as soon as early next year, while Trending Topics describes the IPO as shelved for now, noting that Sam Altman said in September that going public would be "an ill-advised moment" given safety questions, and that CFO Sarah Friar had earlier pointed staff to 2027.
The scale of what the money has to cover explains the nervousness. Figures cited by Trending Topics from FT and Reuters reporting put OpenAI's annualised revenue run rate at roughly $50bn at the end of September, against earlier investor guidance of $68bn to $70bn, with the gap attributed mainly to revenue flowing through cloud partners that OpenAI does not fully book. The same reporting puts expected total cash burn through 2030 at $278bn, alongside a compute contract with Oracle worth about $300bn. These are media figures from unaudited sources, not company disclosures, and should be read as the range investors are arguing about rather than settled accounts.
None of this says OpenAI is short of money. A $20bn bridge round at a $1.25tn valuation would be among the largest private financings ever, and the prior round raised $122bn in March at an $852bn valuation. What has changed is the direction of travel: for the first time in this cycle, a frontier lab has had to take a lower number than it asked for, and the reason given across every account is investor doubt about returns on AI capital spending. OpenAI did not comment to Investing.com, CNBC or the FT, and Sequoia, Founders Fund and SoftBank declined to comment to the FT.
Why it mattersInvestor appetite is what has kept frontier AI cheap for the people building on top of it. A lab taking $10bn less than it asked for is the clearest signal yet that the subsidy behind low API prices and generous subscription limits is finite.
👀 What to watch
The first close figure and whether it lands below $20bn, Anthropic's mid-November listing and the valuation it achieves, and whether OpenAI files to list in early 2027 as the Journal reported.
3 opportunities from this story
Vendor-dependency review for products built on one AI provider
Plenty of small products are one price change or one deprecated model away from a broken margin. Sell a short review that maps every AI dependency in a client's product, prices the three most likely adverse changes, and specifies the swap that would keep the product running. The output is a two-page risk note a founder can show a client or an investor. Funding wobbles at the top of the stack make this an easy conversation to start this month.
- Best for
- Technical consultants, fractional CTOs, senior developers who freelance
- First step this week
- Pick one product you know well and write the two-page note for it unpaid, then use it as your sample.
Open the full playbook
Launch steps
- List every model, API and feature the product relies on, including hidden ones in libraries
- For each, name a concrete substitute and estimate the work to switch
- Price three scenarios: a price rise, a model retirement, and a rate-limit change
- Write the note in plain language for a non-technical founder
- Offer the fallback implementation as a separate quote
Tools
Risks
You are selling a document, which is easy to dismiss as a slide deck. Keep it concrete by naming real substitutes and real hours, and offer to implement one fallback as proof it works.
Bootstrapped niche AI tools priced to survive without funding
If capital is harder to raise, the advantage shifts to small products that are profitable at a few hundred paying users. Pick one narrow, unglamorous workflow in a market you understand, build the smallest tool that does it with a cheap model, and price it so that one hundred customers cover the bills. Local-language and local-compliance niches are good hunting ground precisely because they are too small for a funded startup to bother with.
- Best for
- Indie builders, developers with a day job, domain experts who can code a little
- First step this week
- Write down three workflows you have personally been paid to do manually, and pick the one with the most repetitive input.
Open the full playbook
Launch steps
- Interview ten people who do the workflow by hand and record what they would pay
- Build the thinnest version that handles the single most common case
- Price with your own model costs plus a margin, and cap usage per tier
- Charge from the first customer rather than running a free beta
- Add a second case only after the first is reliably paid for
Tools
Risks
Niche markets are small by definition, so validate willingness to pay before building. Model costs can move against you, so keep a usage cap in every tier and review pricing quarterly.
Plain-English AI spending briefings for local business owners
Most owners of small and mid-sized firms are hearing about an AI bubble and have no way to judge whether it affects them. Produce a short monthly briefing that translates the funding and pricing news into what it means for their tool budget and their vendors, and sell it to a chamber of commerce, an industry association or a handful of firms directly. It is a writing product, so the cost of starting is your time, and it builds the credibility that leads to consulting work.
- Best for
- Writers, analysts, consultants building a local reputation
- First step this week
- Write one issue this week covering this month's AI pricing changes and send it free to twenty business owners you know.
Open the full playbook
Launch steps
- Define the audience narrowly: one city, one sector, or one association
- Fix a format of two pages: what changed, what it costs you, what to do
- Publish three free issues to prove consistency
- Pitch one association for a bulk subscription for its members
- Add a paid consultation slot at the bottom of every issue
Tools
Risks
Commentary is a crowded market and trust takes months to build, so cite sources and never predict markets. Treat the briefing as lead generation for paid work rather than as the main revenue line.
Aramco buys 49% of Saudi AI firm HUMAIN, which plans 3 gigawatts of data centres by 2030
Saudi Arabia's Public Investment Fund and Aramco said on 1 November that they had closed a deal giving Aramco 49% of the AI company HUMAIN, in exchange for about SAR 9.9bn ($2.6bn) in cash during 2026 plus AI, compute and digital assets. HUMAIN plans to deploy 3 gigawatts of AI data-centre capacity by 2030.
Aramco and Saudi Arabia's Public Investment Fund announced on 1 November, from Dhahran, that they had completed the transaction giving the oil company a minority stake in HUMAIN. PIF retains 51% and Aramco takes 49%, held through a wholly owned subsidiary, so the fund keeps control. The consideration, in the companies' words, is "approximately SAR 9.9 billion ($2.6 billion) in cash contributions in 2026 from Aramco to HUMAIN, in addition to AI, compute, and digital assets". The release does not itemise which assets transfer, and gives no valuation for HUMAIN as a whole.
HUMAIN itself is barely eighteen months old. PIF launched it in May 2025, and Aramco's release describes it as building "full-stack AI capabilities" across four areas: next-generation data centres; infrastructure and cloud platforms; AI models, including ALLaM, which the release calls one of the world's most powerful multimodal Arabic large language models; and applied AI solutions. The stated ambition is for Saudi Arabia "to become one of the leading exporters of AI globally". The headline capacity number is the deployment of 3 gigawatts of AI data-centre capacity by 2030, which the release ties directly to what this transaction is meant to enable.
Three gigawatts is a serious figure rather than a slogan, and worth putting in context. It is comparable in scale to the largest announced AI campuses anywhere, and in a country where land, sunshine and gas are cheap but water is not, the constraint is likely to be cooling and grid build rather than capital. The release gives no current capacity figure, names no chip supplier, and does not say how the 3GW is phased between now and 2030. Those omissions matter: capacity targets announced four years out have a poor record of arriving on schedule, and nothing here commits HUMAIN to a particular hardware vendor or a particular year.
For Aramco, the logic is stated as industrial rather than financial. Amin H. Nasser, the company's president and chief executive, said "the HUMAIN deal is very important to Aramco" and noted that "since 2023, Aramco has achieved $11.3 billion in Technology Realized Value, with a significant portion coming from AI solutions" — a company-defined metric, not an audited revenue line. Yazeed A. Al-Humied, PIF's deputy governor and head of MENA investments, framed the combination as creating "a powerful synergy that fuels talent growth, technological advancement, and intellectual property creation". The expectation on Aramco's side is to apply HUMAIN's infrastructure to industrial use cases in its own operations.
Read next to OpenAI's smaller funding round, the direction is striking. Private AI capital in the United States is becoming more cautious at exactly the moment Gulf state capital is becoming less so, and the same UAE money reported as the likely anchor of OpenAI's round is part of the same pattern. The practical effect for people outside the region is that a growing share of new AI compute, and the jobs that come with building and running it, will sit in the Gulf. The open questions are who supplies the chips, what the power and water footprint looks like, and whether capacity aimed at export markets finds customers outside Saudi Arabia.
Why it mattersState capital is now funding AI compute at a scale private investors are starting to hesitate at, and a 3GW target aimed at export shifts where the next wave of AI infrastructure work will physically sit.
👀 What to watch
Which chip and cooling suppliers HUMAIN names, whether any phasing detail for the 3GW target appears, and whether ALLaM is offered commercially outside Saudi Arabia.
3 opportunities from this story
Remote specialist services into Gulf AI infrastructure projects
A 3GW build programme needs far more than chip engineers: documentation, monitoring dashboards, data-centre operations runbooks, procurement admin and multilingual support. Position yourself for one narrow, remotely deliverable slice of that, and target the contractors and integrators around the project rather than the headline company. African and Asian freelancers already serve Gulf clients in adjacent fields, and timezone overlap with East Africa is close to perfect.
- Best for
- DevOps engineers, technical writers, data-centre operations specialists, project coordinators
- First step this week
- Identify five engineering or integration firms with announced Saudi data-centre work and find the exact job titles they are hiring remotely.
Open the full playbook
Launch steps
- Pick one deliverable you can prove: runbooks, monitoring, or compliance documentation
- Build a portfolio sample using a public reference architecture rather than client data
- Register on the Gulf contracting and freelance platforms the integrators actually use
- Approach subcontractors and integrators, not the principal
- Check visa, tax and invoicing requirements before quoting cross-border
Tools
Risks
Cross-border contracting carries payment and tax risk, and large infrastructure programmes slip. Insist on milestone payments, keep other clients running, and do not reorganise your business around one unstarted project.
Industrial AI pilots for local manufacturers and utilities
Aramco's stated reason for buying into HUMAIN is to apply AI to industrial operations. The same playbook works at a much smaller scale for factories, mills, bottlers, water utilities and transport firms in your own market, which generate sensor and maintenance data nobody analyses. Sell a short, scoped pilot on one machine or one line: predictive maintenance alerts, quality-check photographs, or fuel and energy anomaly detection, with a clear measure of whether it worked.
- Best for
- Data analysts, IoT developers, engineers with factory-floor experience
- First step this week
- Approach one local plant manager and ask what single recurring breakdown or reject costs them the most money each month.
Open the full playbook
Launch steps
- Scope the pilot to one machine, one failure mode and one measurable outcome
- Audit what data already exists, including paper maintenance logs
- Instrument the minimum needed, often a single sensor or a phone camera
- Run four to eight weeks with a baseline and a clear success threshold
- Present the result honestly, including a null result, and quote the rollout
Tools
Risks
Industrial data is messy and sales cycles are long, so you may spend weeks on cleaning before any model work. Charge for the pilot, scope it tightly, and be willing to report that the data was insufficient.
Non-English AI evaluation and localisation work
ALLaM is a reminder that serious money is going into models for languages that English-first benchmarks ignore, and that creates demand for people who can judge output quality in those languages. Build a small, reliable evaluation service: native-speaker review panels, test sets for a specific language or dialect, and reports a model team can act on. The same capability sells to any company shipping an AI product into a market whose language it cannot check itself.
- Best for
- Linguists, translators, researchers, anyone with a native-speaker network
- First step this week
- Build a 100-item test set in one language you know well, covering translation, summarising and refusal cases, and run it against two public models.
Open the full playbook
Launch steps
- Choose one language or dialect and define what good output means for three task types
- Recruit three to five native reviewers and agree a scoring rubric
- Publish a short public report comparing two or three models on your test set
- Pitch model teams, localisation agencies and companies entering that market
- Keep the test set private once clients start paying, and version it
Tools
Risks
Reviewer quality varies and inter-rater disagreement can sink a report, so pilot the rubric before scaling. Rates for annotation work are under pressure, so compete on expertise and reporting rather than on price.
Get these as a PDF every 3 days
Free. One email every 3 days. Unsubscribe any time.