This website uses cookies

Read our Privacy policy and Terms of use for more information.

Australia's biggest float in years asked for nearly three times the last private price.

The book closed one day and the deal was pulled the next. The company had more than nine hundred megawatts planned and forty-six running. Now it is going back to private money. A rival raised up to four billion dollars privately that week.

Private money still pays that price.

Public money just said no.

I'm Ben Baldieri. Every week, I break down what's moving in GPU compute, AI infrastructure, and the data centres that power it all.

Here's what's inside this week:

Let's get into it.

Firmus Pulls Its A$7.1bn IPO and Goes Back to Its Backers

Australia's biggest float in years died on a thin book, and Firmus is going back to the investors who valued it at a third of the float price.

Firmus shelved its IPO on 9 October, a day after its institutional bookbuild closed, Reuters reported. The board said going ahead "was not in the best interests of the company and its shareholders" and blamed market volatility. Bloomberg had reported a day earlier that Firmus was weighing a private round with current investors and new ones. It had set out to raise A$7.1 billion at A$11 a share. Reuters put the equity value at that price at US$30.6 billion. That is nearly triple the August round, which raised US$2 billion from NVIDIA, Coatue, Blackstone and Jane Street, and about US$30 billion of debt sits on top. Escrow terms would have let existing holders sell more than half their stock from day one. The book never got there. Reports had the price sliding to between A$8 and A$9 before the deal was pulled, and DCD reported days earlier that CDC Data Centres had ended Project Southgate after 42MW of a planned 1.6GW. Jun Bei Liu of Ten Cap called the pulled deal "an important reality check for the AI investment boom".

Why this matters:

OpenAI Names Its Own Price: $1.4tn, With No Lead Investor

OpenAI wants $30 billion at a price no lead investor has tested.

OpenAI is talking to Abu Dhabi's MGX and BlackRock about raising $30 billion at a $1.4 trillion pre-money valuation, Bloomberg reported on 5 October. The round itself is not new (#126). The names and the missing lead are. MGX is putting together a syndicate of UAE funds worth up to $10 billion. The University of California endowment, Thrive Capital and Andreessen Horowitz are also named. No investor is leading the round. OpenAI set the valuation itself. Its last round raised $122 billion in March at $852 billion, and Bloomberg reports the listing has been pushed back. The FT has since reported that OpenAI's annualised revenue is about $20 billion lower than previously signalled, Reuters reports.

Why this matters:

  • Without a lead, nobody tested the price. The investors are paying the number OpenAI chose.

  • The money is Gulf infrastructure capital. MGX already runs a $30 billion infrastructure venture with BlackRock, Microsoft and NVIDIA. It is also a partner in Mistral's 1.4GW campus outside Paris.

  • The two largest labs now fund compute in opposite ways. Anthropic borrows from its chip supplier. OpenAI sells equity to sovereign funds at a price it picks.

Lambda Lines Up $4bn in Private as Anthropic Fills 70% of Its Backlog

Lambda is raising privately in the same week public buyers turned Firmus down.

Coatue and Blackstone are leading a raise that could reach $4 billion, the Wall Street Journal reported on 6 October. It values Lambda at $14.5 billion before the new money and comes ahead of a planned listing. Lambda's backlog jumped from $15 billion in June to $50 billion in September. One Anthropic contract makes up $35 billion of it.

Why this matters:

  • Lambda's valuation leans on one customer. Anthropic is 70% of the backlog, and Anthropic borrows from its chip supplier (#126).

  • The backers overlap. Coatue and Blackstone both backed Firmus, Reuters reports, and are now funding Lambda before it lists rather than after.

  • Private money still prices neoclouds above public money. Lambda is raising before its listing, and Firmus went back to private money after its listing failed.

Mistral Trained a Trillion Parameters on 3,800 GPUs. Aleph Alpha Used a Thousand.

Two European labs shipped open models in one week, and one of them is about to trade as Cohere.

Mistral put Large 4 into public preview on 6 October. It is a mixture-of-experts model with 1 trillion parameters, 49 billion of them active per token. It reads text and images in more than 160 languages. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres. Pierre Stock, its vice-president of science, told TechCrunch that is two to three times less compute than its Chinese rivals use. TNW puts the run at about 10MW over two months. Pricing is $1.36 per million input tokens and $4.18 per million output tokens. Weights are due by the end of October, under a licence Mistral has not named. Every benchmark in the launch post is Mistral's own. Days earlier, Aleph Alpha released Kolibri, a 78 billion parameter model with about 3.5 billion active, under Apache 2.0. Around a fifth of its training data is German. Finnish neocloud Verda, which hosted the run, says it took over a thousand NVIDIA Blackwell GPUs running around the clock for months. Aleph Alpha signed a definitive merger with Cohere in September. Reports value the combined group at about $20 billion, and it will trade as Cohere.

Why this matters:

  • Europe is selling efficiency while its scale is still under construction. Mistral's run drew about 10MW, against the 1.4GW campus planned outside Paris. Anthropic's leaked draft showed $518 billion of compute commitments (#126).

  • Both runs used European-operated capacity. Mistral trains in its own halls and now rents the spare as Mistral Compute. Aleph Alpha used Verda, which raised $189 million in September (#125).

  • Sovereignty has an owner. Kolibri is a German model, its maker is about to trade as Cohere, and Schwarz Group has committed €500 million to lead Cohere's next round. Neither model has independent benchmarks yet.

NVIDIA and Microsoft Launch RTX Spark PCs With 128GB for Local AI

The local AI box now comes from the company that sells the cloud GPUs.

NVIDIA and Microsoft launched RTX Spark PCs on 7 October. The chip pairs a Blackwell RTX GPU with 20 Grace CPU cores and takes up to 128GB of unified memory. NVIDIA rates it at one petaflop of FP4. NVIDIA says it runs Qwen 3.8 Flash Next, a 125 billion parameter model, with no cloud connection. Acer, ASUS, Dell, HP, Lenovo, MSI, Gigabyte and Microsoft's Surface all have machines. Laptops go on sale on 16 October. Microsoft prices the Surface Laptop Ultra from $2,599 and its RTX Spark Dev Box at $5,999. Ghost had left stealth two days earlier with an $11 million seed round led by Andreessen Horowitz. Its $3,499 Core box has a 24GB RTX PRO 4000 Blackwell GPU, ships with Qwen 3.8 27B and Gemma 4 31B, and starts shipping on 31 October.

Why this matters:

Amazon Drops Its NDAs as Towns Kill $260bn of Data Centres

Amazon dropped its NDAs a day after pledging $1 billion to the towns that host its data centres.

AWS chief executive Matt Garman said Amazon has stopped using non-disclosure agreements with government agencies when it seeks data centre approvals, TechCrunch reported on 3 October. He disputed the main complaints. He said data centres use 0.5% of US industrial water, that higher power bills come from ageing grids, and that backup generators sit idle 99.9% of the time. A day earlier Amazon pledged another $1 billion over five years to the communities that host its sites. Local opposition has killed at least $260 billion of data centre projects this year, Heatmap reports. Half of that went in the last quarter.

Why this matters:

  • Secrecy was the complaint that stuck. Boyd County residents learned of TeraWulf's Kentucky campus months after their officials did, Bisnow reported in June, and NDAs took the blame.

  • The generator line is the weakest of Garman's four. New Jersey fined a site $1.07 million last week for running 62 gas generators without permits (#126).

  • $260 billion of cancelled projects is the price of losing local trust. $1 billion over five years is what Amazon is paying to win it back.

TeraWulf Booked 1GW in Kentucky and Has to Borrow to Guarantee It

A former bitcoin miner booked a gigawatt in eastern Kentucky, and the utility wants collateral first.

TeraWulf said in an 8-K filed on 5 October that Kentucky Power, an AEP subsidiary, will supply its Muskie Data Campus with 1GW, up from 500MW. The amended agreements were signed on 1 October. Phase one is due to ramp in 2028, and phase two moves forward a year to 2029. Both need approval from the Kentucky Public Service Commission. TeraWulf plans to upsize its senior secured revolver and raise more secured debt to back the letters of credit the utility requires. It had 102MW of revenue-generating capacity at Lake Mariner, New York, after finishing a building in July. Anthropic leases its Hawesville campus, also in Kentucky, for about 401MW over 20 years. TeraWulf values that deal at about $19 billion.

Why this matters:

  • Contracted power now runs at more than ten times operating load. 102MW is running, and 1.4GW is booked across two Kentucky campuses alone.

  • Utilities want collateral before they build. TeraWulf has to borrow to post the letters of credit Kentucky Power requires.

  • Anthropic's $518 billion of commitments (#126) lands on sites like these. A former miner in Kentucky is part of that number.

Everything Else

Reply

Avatar

or to participate