OpenAI said on 11 September that it rewrote the storage system behind ChatGPT, called Habitat, from Python into Rust, taking peak throughput from more than 20 million to 70 million requests every second while holding more than 500 petabytes of data across almost 40 regions. The new version is six times more efficient on processing power and fifteen times on memory, and now serves 95 percent of production requests. OpenAI says ChatGPT reaches 1 billion people each week, and that the rewrite was done in one quarter by two engineers using its own Codex and GPT-5.5. This changes nothing about what you can ask ChatGPT, but it is part of why the product stays responsive while its user base keeps growing, and it is a rare public number for how much of a rewrite this size a lab now hands to its own models.
Canadian AI company Cohere reportedly in talks to raise $2 to $3 billion at a $20 billion valuation
· Band Money · $2-3bn at $20bn, reported · Reported
Cohere, a Canadian AI company that sells its models mainly to businesses and governments, is in advanced talks to raise between 2 and 3 billion dollars at a valuation of 20 billion dollars, according to a Globe and Mail report carried by PYMNTS and attributed to unnamed sources. Neither Cohere nor any named investor has confirmed a figure. If the round closes at the reported size it would be a large step up for the company, but until someone puts their name to a number these should be read as reported rather than settled.
SpaceX signed a new AI computing deal worth 1.1 billion dollars a month, about 13 billion dollars a year, according to comments its chief financial officer Bret Johnsen made at a Goldman Sachs conference, reported by Yahoo Finance. The customer has not been named, the deal was signed earlier in September and payments begin in December. No company document confirming the terms could be found, so this runs as Reported. The same account says SpaceX already hosts AI computing for Anthropic at 1.25 billion dollars a month through May 2029 and for Google at 920 million dollars a month through June 2029, and that this is a separate new arrangement. For most people nothing changes today; this is a corporate financial disclosure about infrastructure.
Microsoft is reported to be planning to more than triple its data centre capacity to about 38 gigawatts by 2032
· L3 Iron and cloud · about 38 GW by 2032, from about 12 GW · Reported
Bloomberg News reported, citing people it did not name, that Microsoft plans to more than triple its data centre capacity to about 38 gigawatts by 2032, up from about 12 gigawatts today. We read this through a trade outlet's account of Bloomberg's reporting rather than Bloomberg itself, which is why it runs as Reported. Microsoft has not confirmed the target, and the account says the company did not immediately respond to a request for comment. If accurate it would be one of the largest capacity expansions by any single cloud company. Nothing changes today; this is an unconfirmed report about long term planning.
Reuters, citing six people it did not name, reported that the United Arab Emirates is redesigning a planned 5 gigawatt AI data centre project, one of the biggest outside the United States, after last year's Iran war prompted a rethink of how and where critical infrastructure should be built. No UAE government body or company has confirmed this account on the record, so it runs here as reporting rather than a confirmed fact. If accurate, it points to AI data centres increasingly being treated as strategic, security-sensitive infrastructure in the Gulf, which could slow down or reshape how quickly that capacity gets built. Nothing changes today for anything already running.
Together AI, a company that rents out computing power for building and running AI models, lowered the price of fine-tuning, the process of further training an existing model on a company's own data, by 30 to 70 percent across most of the models it supports. For one model, OpenAI's open-weight gpt-oss-20b, the standard training price fell from 1.50 dollars to 0.40 dollars per million tokens processed. Together AI says the reduction comes from its own optimisation of the systems that do this training, not from a change in what the underlying chips cost. For a software team or startup that fine-tunes models to fit its own product, this is a direct reduction in one of the recurring costs of building an AI product, though it applies to training a model rather than running it once it is finished.
Australia's public broadcaster ABC reported that 27 data centre projects already approved but not yet built, with a combined capacity of at least 3.38 gigawatts, will not have to meet new water and power standards the government introduced earlier this year, because the rules only apply to projects approved from now on. Prime Minister Anthony Albanese told ABC Radio Sydney, "You can't retrofit. But as well, many of the centres which are under construction tend to be smaller ones," explaining why existing approvals would not be held to the new requirements. As of March 2026 Australia had 162 data centres running with a combined 1.5 gigawatts of capacity, so the exempted projects alone are more than double what the country already operates. For Australians living near one of these 27 sites, the facility being built nearby will not be held to the newer environmental standards that future projects will have to meet.
Aperture AC agrees to buy Atlantic HPC Group in a $150 million deal to take it public
· L1 Power and land · $150M, 51 MW running of a 121 MW pipeline · Confirmed
Aperture AC, a blank-cheque company already listed on Nasdaq, agreed to acquire Atlantic HPC Group, a high-performance computing operator, in a deal that values Atlantic at 150 million dollars before the merger. Atlantic has 51 megawatts of computing capacity running across facilities in Oklahoma, Arkansas and Ohio, out of a development pipeline of 121 megawatts across six sites. This is a small deal by the standards of the AI data centre boom, well under the billion-dollar mark of the largest ones, and the merger still has to close. For anyone tracking how smaller computing operators raise money, this is one more example of a company using a blank-cheque merger rather than a traditional listing or a private round.
Australian data centre operator NextDC launches a 1.1 billion Australian dollar convertible note offering
· Band Money · AU$1.1bn (about US$796m) · Reported
NextDC, an Australian data centre operator, is raising 1.1 billion Australian dollars, about 796 million US dollars, by selling convertible notes, bonds that lenders can later swap for company shares at a fixed price. It is the third time in a little over four months that NextDC has sought fresh capital, following a debt facility in July and other commitments in May. For anyone watching how AI data centre construction gets financed, this is another example of debt, not just equity or cash on hand, funding the buildout.
Pennsylvania's utility regulator voted unanimously on 11 September to approve two motions, Utility Dive reports. One directs its staff to propose updates to the state's rules for emergency curtailment of electricity use and to organise a technical conference on how costs for data centres are allocated. The other starts a working group on improving how the commission sets rates, with a focus on transparency and affordability. The commission cited PJM's July 2026 capacity auction, which it said identified a system-wide reliability shortfall of 6,831 megawatts for 2028 to 2029, and an August report projecting 18.56 percent industrial load growth in the state over five years. For Pennsylvania electricity customers, this starts a process that could decide who pays for grid upgrades that very large users need.
Oklo, a startup building small nuclear reactors partly aimed at powering AI data centres, filed paperwork authorising itself to sell up to 1 billion dollars of its own stock over time, in a program run through a group of ten banks. This is a ceiling on how much Oklo could sell, not money it has raised yet; how much it actually sells, and when, is up to the company. For anyone who owns Oklo stock, this is a tool the company can use to raise cash as needed, which can dilute existing shareholders when shares are actually sold.
Japan's Sakana AI launches two new models, and prices Fugu Max well below Western rivals
· L4 Models · $2 per 1M input, $6 per 1M output · Confirmed
Sakana AI, a Japanese AI lab, released two new models, Fugu Max and Fugu Ultra v2, with Fugu Max priced at 2 dollars per million input tokens and 6 dollars per million output tokens for developers who build on it through the company's API. Sakana says that output price is 40 to 60 percent lower than comparable models from Anthropic, OpenAI and Moonshot, and that its models score best or near best on several of its own benchmark tests against those rivals. These are Sakana's own comparisons on its own chosen tests, not independently verified. For anyone building software on top of AI models, this is more evidence that competition on price, not just capability, is intensifying beyond the handful of US labs that have dominated headlines so far.
Korea's semiconductor exports jump 270 percent in the first ten days of September
· L2 Silicon · $16.483bn semiconductor exports, +270.1% year over year · Reported
Korea's customs service says the country exported about $35 billion worth of goods in the first ten days of September, a record for the period and up 82.6 percent from a year earlier, according to a Korean financial newspaper's report of the announcement. Semiconductor exports alone reached $16.483 billion, up 270.1 percent, and now make up 47.1 percent of all Korean exports, a share 23.9 percentage points higher than a year ago. This is early, ten day data, not a full month. For an ordinary person nothing changes today; this is an early read on Korea's export economy, not a price.