AI Markets: Microsoft thinks big, Oracle reassures, Anthropic eyes $2 trillion

Between Microsoft’s 38-gigawatt plan, Oracle’s reassuring numbers and the resignation rattling Anthropic’s pre-IPO run, money keeps pouring into AI – but every player now has to prove the numbers hold up. While SpaceX signs new compute contracts with a surprisingly flexible structure, investor Michael Burry is quietly scaling back his bearish bets on the sector.

Microsoft wants to more than triple its AI data centers by 2032

According to Bloomberg, Microsoft is drawing up an internal plan to grow its global data center capacity to more than 38 gigawatts by 2032, up from about 12 GW today. It’s a bet worth several hundred billion dollars to finally escape the compute shortage that has been costing it customers.

Key points:

On September 10, Bloomberg revealed an internal plan targeting more than 38 GW of data center capacity for Microsoft by 2032, versus about 12 GW today — more than a tripling.
Of the current 12 GW, only about 2 GW is dedicated to AI chips; that share is set to climb to nearly a third of the 38 GW targeted for 2032.
Several Asian outlets reported the same day that Microsoft has had to turn away cloud and AI customers for lack of available servers (“lost customers due to server shortage”).
Bloomberg’s sources note that these internal projections are still subject to change: construction timelines, regulatory constraints, technological breakthroughs.
Nvidia, a potential beneficiary through additional chip orders, rose 0.5% after hours on that news alone.

A compute shortage that costs customers, not just a marketing plan

The 38 GW figure is not an official target announced by Microsoft: it’s an internal projection reported by Bloomberg’s sources, and should be read as such. But it confirms something already documented this year: major cloud and AI customers (enterprises, but also labs like Anthropic or OpenAI) keep running into a capacity wall at every hyperscaler. Amazon already raised its 2026 capital spending budget to $220 billion this summer for the same reason, and Oracle keeps adding data centers to fulfill an AI order backlog that now tops $660 billion (see below). Microsoft, which long leased capacity from third parties rather than building massively on its own, appears to be shifting here to a far more capital-intensive strategy — at the risk of further fueling the debate over the financial sustainability of this infrastructure race.

What it means for a European investor

Microsoft is listed on the Nasdaq and is not directly eligible for the PEA: only an ordinary brokerage account lets you hold its shares individually. For exposure inside a PEA, you need a synthetic-replication fund such as the Amundi PEA Nasdaq-100 UCITS ETF, which holds US stocks within the French tax wrapper. A plan with a 2032 horizon is, by nature, a statement of intent: nothing guarantees it will be executed at the pace, or for the amount, described by Bloomberg’s sources.

Sources:

Oracle eases debt worries with results that beat expectations

After months of anxiety over its debt load and cash burn, Oracle reported quarterly results on September 10 that beat expectations on nearly every metric. Wall Street immediately erased part of the skepticism that had built up since its credit rating was downgraded in July.

Key points:

Total revenue for fiscal Q1 2027 (June-August): $19.3 billion, up 30% year over year; cloud infrastructure (OCI) jumped 121% to $7.4 billion.
The order backlog (RPO) reached $664 billion after more than $30 billion in new AI contracts signed during the quarter — above the $640 billion analysts expected, according to Reuters.
Cash burn was much lower than feared: negative free cash flow of $5.4 billion versus an expected -$9.56 billion, alongside record operating cash flow of $23 billion for the quarter.
Guidance for fiscal 2027 was raised to at least $90 billion in revenue and $8.10 in adjusted earnings per share; the stock climbed 7 to 8% after hours.

The number that reframes the debate: less cash burned than forecast

In early September we covered the concerns around Oracle: a new wave of layoffs meant to plug an estimated $20 billion funding gap in its AI data center buildout, after free cash flow of -$23.7 billion for full fiscal 2026 and total debt close to $129 billion. This quarter doesn’t close the file — debt remains high and S&P downgraded Oracle’s rating in July — but it does bring a new fact to the table: the quarter’s cash burn was half what analysts expected, and the order backlog keeps growing faster than feared. Enough to temper, without settling, the debate between believers in an AI “super-cycle” and those who see a credit-fueled spending bubble.

European investor angle

Oracle (ORCL) is listed on the NYSE and remains outside the PEA: only an ordinary brokerage account, or an ETF holding US stocks such as the Amundi PEA Nasdaq-100 UCITS ETF, provides exposure from within a French tax wrapper. Note that despite this one-session rebound, the stock was still down more than 20% since the start of 2026 before the release: one reassuring quarter doesn’t settle, on its own, the debate over the sustainability of the sector’s debt.

Sources:

SpaceX signs a $13 billion-a-year AI contract… with a 6-month exit door

At a Goldman Sachs conference on September 10, SpaceX confirmed a new AI compute hosting contract worth more than $13 billion on an annualized basis. But the deal’s structure, designed to let SpaceX take back its capacity at any time, tempers the promise of recurring revenue. The same day, the company also found itself in a Texas court trying to block the release of documents about its Terafab plant.

Key points:

CFO Bret Johnsen announced a new AI compute hosting contract worth about $1.11 billion per month starting December 1, or roughly $13.3 billion annualized — client not identified.
The contract follows the same structure as SpaceX’s previous compute deals (with Google or Anthropic): an initial 90-day commitment followed by a 90-day exit clause, meaning about 6 months before either party can walk away — so this is not $13 billion guaranteed over a year.
Johnsen reaffirmed the target of $100 billion in annualized revenue (ARR) by year-end, including this new contract.
Starship is set to fly its first paid commercial mission in the coming weeks, carrying production Starlink satellites.
Separately, on September 8, SpaceX sued Grimes County, Texas, and the Texas attorney general to block the release of tax documents related to Terafab, its semiconductor and AI hardware plant run as a joint venture with Tesla ($16.8 billion invested, 35-year tax exemption).

Why the “6-month clause” matters if you follow the stock

Since this summer, SpaceX has been building a narrative of AI revenue that would soon exceed Starlink and launches combined, with contracts already signed with Google ($920 million/month) and Anthropic ($1.25 billion/month). This new deal joins the list and brings the company closer to its $100 billion ARR target. But an ARR (“annual recurring revenue”) calculated by annualizing a single month of billing only makes sense if the contract actually lasts a year — and the 90-day exit clauses, disclosed by Johnsen himself, let both SpaceX and its client pull out well before then. That’s a factual nuance, not a verdict on the strength of the model: it simply explains why these figures deserve to be read as activity indicators rather than as revenue guaranteed over twelve months.

A legal fight at odds with the growth story

The lawsuit against Grimes County is a separate development from the already-known Terafab story: this time SpaceX is trying to block the release of its tax exemption file, citing competitive harm. The episode illustrates the growing local friction the company faces as it expands its industrial footprint across the US (Terafab, the Bastrop turbine foundry, the Louisiana spaceport), already covered in our previous editions.

European investor angle

SpaceX (SPCX) is listed on the Nasdaq and remains outside the PEA (ordinary brokerage account only); its growing weight in the Nasdaq-100 index does, however, give indirect exposure via the Amundi PEA Nasdaq-100 UCITS ETF, the only PEA-eligible tracker of that index.

Sources:

Anthropic: a high-profile resignation crashes into the sprint toward a $2 trillion IPO

As Anthropic aims to publicly file its prospectus as early as late September for an IPO floated at a valuation of up to $2 trillion, a researcher has resigned, forfeiting his unvested shares, to publicly sound the alarm on AI’s existential risks — reigniting, right in the middle of IPO prep, the debate over how solid the generative AI narrative really is.

Key points:

Jacob Coxon, a researcher at Anthropic, resigned in early September, giving up his unvested stock options and claiming that the teams building AI “sincerely believe” it could end humanity by the end of the decade.
His post on X has topped 115 million views; he says he “no longer has anything to gain from pumping up Anthropic’s valuation” since he left before his shares vested.
Elon Musk publicly mocked the exit, calling it a “psy-op”; Anthropic has not commented publicly.
Meanwhile, several investors quoted in the financial press are now floating an IPO valuation that could reach $2 trillion, or even $3 trillion, for Anthropic, with a prospectus filing targeted for late September and a listing in mid-October — neither officially confirmed by the company.
That level would be more than double the $965 billion valuation from its May funding round, and would top the current record of $1,770 billion set by SpaceX at its June IPO.

Timing that raises questions, not proof in itself

This departure isn’t an isolated case: another member of Anthropic’s safety team resigned earlier this year for similar reasons. But its viral reach (115 million views) and its timing, just weeks before an IPO meant precisely to convince the market of the strength and safety of Anthropic’s business model, set it apart. It’s worth separating the facts (a resignation, a real financial sacrifice, a massive audience) from Coxon’s own interpretation (an existential risk he presents as a shared internal conclusion, without documenting it in detail): Fortune notes that his warning sidesteps the hardest question of all, namely what to actually do about that risk. According to earlier reports, Anthropic’s prospectus already lists growing American public backlash against AI as an official risk factor.

European investor angle

Anthropic remains a private, unlisted company for now. If the IPO goes ahead on the Nasdaq as planned, it won’t be directly eligible for the PEA (US stock, ordinary brokerage account only); ETF exposure would only be possible after its eventual inclusion in a broad index, several months after listing. Nothing in this story makes it possible at this stage to predict the final IPO price.

Sources:

Michael Burry eases off his bearish bets as Palantir signs with Nvidia at 55 times sales

The investor famous for calling the subprime crisis has quietly closed out his put options on Nvidia and Palantir expiring in December, without rolling them over. The same day, Palantir announced a partnership with Nvidia on “sovereign” AI — greeted by a slight dip in the stock, as investors questioned a valuation that no longer leaves room for “good enough.”

Key points:

In September, Michael Burry (Scion Asset Management) closed out all of his put options on Nvidia and Palantir expiring in December 2026, without rolling them to a later expiration.
He does, however, keep puts expiring in 2027 on Palantir and on the Invesco QQQ ETF (which tracks the Nasdaq-100): so he isn’t fully abandoning his bearish thesis on AI tech.
Burry says he wants to keep cash on hand for “what looks to be an interesting fall,” rather than signaling a change of conviction; he also trimmed several of his long positions (Lululemon, Molina Healthcare, MercadoLibre).
On the same day, September 10, Nvidia and Palantir announced a partnership to deploy a “sovereign” AI stack across supply chains (agriculture, manufacturing, pharma, public sector), combining Nvidia’s Nemotron models with Palantir’s Foundry/AIP platform.
Despite the announcement, Palantir stock fell about 0.8% that day: measured against its annualized Q2 revenue (about $7.76 billion), its market cap of roughly $430 billion works out to nearly 55 times sales.

Two signals that tell different stories

Closing out options that expire in the coming months isn’t the same as a change of conviction: it’s also standard management of time decay (theta) on contracts that lose their value if they aren’t rolled over in time. The fact that Burry is keeping 2027 puts on Palantir and the Nasdaq-100 while trimming his long positions points more to a broad defensive stance than to a bullish bet on AI. Conversely, the lukewarm reception for the Nvidia-Palantir partnership shows that good business news is no longer enough to lift a stock when its valuation (55 times sales) leaves, in the very words of the analysis we reviewed, “little room for merely decent execution.”

European investor angle

Nvidia and Palantir are listed on the Nasdaq and remain outside the PEA for direct holdings (ordinary brokerage account), except for indirect exposure via the Amundi PEA Nasdaq-100 UCITS ETF, which holds both stocks within a French tax wrapper. Above all, this episode shows how hard it is to read a clear signal in the AI valuation debate: an iconic investor scales back his bearish hedges at the very moment a major tech partnership makes the market question the price being paid for its execution.

Sources:

Edition prepared by our monitoring agent (data and sources as of Friday, September 11, 2026), reviewed according to our editorial policy. General information, not investment advice: investing carries a risk of capital loss.