AI Markets: Starship aims for orbit, Anthropic reportedly eyes November

SpaceX is playing for high stakes this Monday with Starship’s first orbital flight, while Anthropic’s IPO may slip to November. On the chip front, Beijing could reopen the door to Nvidia, Micron reports Wednesday and Optimus’s hands are slowing Tesla down.

Starship aims for orbit this Monday: what Flight 14 means for SpaceX stock

SpaceX is attempting Starship’s first orbital flight this Monday, with 26 Starlink V3 satellites on board. Liftoff is scheduled for 2:15 p.m. Paris time, and the outcome was not yet known at the time of writing.

Key points:

  • Liftoff is targeted for September 28 at 7:15 a.m. Texas time (2:15 p.m. in Paris), within a 75-minute window, from Starbase. The FAA issued its license Saturday night, according to NASASpaceFlight.
  • The ship is supposed to reach an orbit of roughly 275 km, circle the Earth about six times over nearly ten hours, then deploy 26 Starlink V3 satellites.
  • The flight was originally scheduled for September 22. SpaceX gave no official reason for the six-day delay.
  • On the market, SPCX shares fell more than 4% on September 23, the day before roughly 328 million shares were unlocked.

What Flight 14 needs to prove

Starship’s first thirteen flights were suborbital. This time, the ship has to stay in orbit, deploy its payload, then reenter and splash down in the Pacific, west of Chile. The first stage, for its part, will attempt a landing at sea.

The satellites are real Starlink V3 units, not a dummy mass. That’s why Yahoo Finance calls this flight Starship’s first revenue-generating mission. According to NASASpaceFlight, each V3 adds about 1 Tbps of capacity to the network, or 26 Tbps for this batch alone. The link to the stock is direct: Starlink has been the group’s only profitable segment since its first quarterly results in August. The flight is still a test, and its outcome is guaranteed neither way.

A stock under pressure before liftoff

Three factors weighed on SPCX this week. The first is the unlock of roughly 328 million shares on September 24. The previous one, on September 9, was followed by a 3.9% drop, according to Yahoo Finance. The second is President Gwynne Shotwell’s sale of 342,170 shares on Tuesday, for about $52.5 million, at prices of roughly $151 to $155, according to an SEC filing reported by Benzinga. The sources consulted do not say why she sold. The third is the flight delay.

According to Yahoo Finance, SPCX lost more than 2% again on Friday. The stock remains above its $135 IPO price, having traded above $150 in mid-September. Analyst Adam Jonas (Morgan Stanley, Buy rating) sees this flight as proof of the commercial viability of the group’s ambitions in AI and orbital computing. That is an opinion.

For a European investor, SPCX is held in a standard brokerage account, not in a PEA. The stock is also part of the Nasdaq-100, which it joined on July 7, and therefore of the ETFs that track it.

Sources:

Anthropic: an IPO reportedly slipping to November, amid Akamai, the Pentagon and the White House

Three new developments on Anthropic since our last edition: an IPO delay reported by the Wall Street Journal, an $11.6 billion cloud contract with Akamai and a court loss against the Pentagon. Donald Trump was also set to have dinner with Dario Amodei on Sunday evening.

Key points:

  • According to the WSJ (September 18), the IPO would move from October to November, so that third-quarter results can be presented first. The article consulted contains no comment from Anthropic.
  • The WSJ mentions a valuation of about $2 trillion, a raise of up to $100 billion and annualized revenue above $110 billion by the end of 2026. None of these figures has been confirmed by Anthropic.
  • Akamai announced an $11.6 billion, seven-year commitment, paired with a warrant that could give Anthropic up to roughly 5% of its equity.
  • On September 25, a federal appeals court upheld, by a 2-1 vote, the designation of Anthropic as a supply chain risk for the Pentagon.

A shifting timeline

Until now, the press had pointed to IPO marketing starting in mid-October and a listing before the midterm elections. The WSJ pushes the whole schedule back a month. It cites the backdrop of Dario Amodei’s public warnings about AI risks, and notes that the company’s advisers believe a slower model release cadence would not weigh much on revenue. The same article says OpenAI is reportedly targeting a $1.2 trillion valuation for its funding round and pushing its own listing to 2027.

The Motley Fool calculates that a $2 trillion valuation on $110 billion in annualized revenue would come to roughly 18 times sales. In our sources, no public prospectus (S-1) had been filed as of the time of writing.

Akamai, the Pentagon, the White House

On Thursday, September 24, Akamai announced an $11.6 billion, seven-year commitment for cloud infrastructure built on conventional processors. The amount could rise to about $20 billion, according to TechCrunch. The warrant covers 7.7 million shares at $111.33: about 2% of the equity would vest with the initial commitment, then about 1% more for each additional $3 billion tranche.

Akamai had a market value of $15.87 billion before the announcement, according to 24/7 Wall St. The commitment therefore amounts to nearly three quarters of that value (Five Minute AI calculation). The stock gained as much as 15% intraday on Friday, while CoreWeave and Cloudflare rose only about 1%. Akamai plans $5.5 billion in capital spending before the first revenue comes in.

On the political front, the Washington appeals court found that the Pentagon had solid grounds to consider the use of Claude in its systems risky. Yet a federal judge in San Francisco had ruled the other designation unlawful the previous month. Two days after the ruling, Trump was set to host Amodei at the White House, their first one-on-one meeting according to Axios. At the time of writing, what was said there was still unknown.

Once listed, Anthropic would be a US stock, and therefore not eligible for direct holding in a PEA.

Sources:

SpaceXAI targets more than 1.2 million Nvidia chips in Memphis by the end of December

On Friday, Elon Musk laid out on X the timeline for his Colossus 2 supercomputer. The number of Nvidia chips in service is set to more than double by year-end, a goal Musk himself says depends on a bit of luck.

Key points:

  • Colossus 2 runs about 550,000 Nvidia chips: 110,000 GB200s and 440,000 GB300s.
  • Three waves of 220,000 GB300s have been announced: next week, in November, then at the end of December “if we get lucky,” according to Musk.
  • The total would then exceed 1.2 million chips. Including Colossus 1, the Memphis site would approach 1.44 million chips, according to WION and Tom’s Hardware.
  • The September 30 deadline in the contract with Google, covering about 110,000 GPUs, falls this week.

A timeline that is still just a plan

These figures come from a post by Musk, and no independent source confirms them. Going from 550,000 to more than 1.2 million chips in three months would multiply the fleet by 2.2 (Five Minute AI calculation). WION notes that the power capacity and cooling for a fleet that size have not been confirmed. Tom’s Hardware writes that SpaceXAI is building a 1.2 GW power plant to supply the whole thing.

The doubt isn’t theoretical. In early September, The Information reported civil engineering problems at the Memphis sites and a leadership shake-up in the data center division. Delivered chips are useless until power and cooling keep up.

Google, Anthropic: compute becomes a product

SpaceX rents this capacity out to third parties. The contract with Google calls for $920 million a month for about 110,000 GPUs, from October 2026 to June 2029, according to DCD. That works out to roughly $11 billion a year (Five Minute AI calculation). If SpaceX doesn’t deliver the promised access by September 30, Google can terminate immediately, or accept the number of GPUs delivered at a reduced rate after a one-month grace period. Anthropic also rents capacity: 220,000 GPUs and 300 MW, according to Tom’s Hardware.

The connection to Nvidia is direct. The Motley Fool describes SpaceX as an “anchor” customer that uses Nvidia chips exclusively. During Nvidia’s August 26 earnings, CFO Colette Kress had already named SpaceX AI among the first customers for Vera chips.

Both Nvidia and SpaceX are held in a standard brokerage account, since neither is eligible for the PEA. Semiconductor UCITS ETFs are the other way in from Europe.

Sources:

Beijing could let Alibaba and ByteDance buy a new Nvidia chip

According to The Information, China is considering allowing groups such as Alibaba and ByteDance to buy Nvidia’s RTX Pro 5500 card. The news sent Chinese chip stocks lower on Monday, including Cambricon and Moore Threads.

Key points:

  • China’s Ministry of Industry and Information Technology asked several companies, including Alibaba and ByteDance, how many units they planned to buy. Some were reportedly told the government plans to approve, according to The Information (September 27).
  • Reuters could not verify the report. Alibaba and ByteDance did not respond to its requests.
  • The RTX Pro 5500 is a Blackwell-generation professional card that Nvidia unveiled earlier this month. Some industry executives believe it falls outside US export restrictions.
  • On Monday, Cambricon lost 5.7%, Moore Threads 6.3% and SMIC 3.7%, according to Investing.com.

Why a workstation card matters

The RTX Pro 5500 packs 84 GB of GDDR7 memory and is designed for server rack configurations, according to Digital Today. It can handle heavy AI workloads, but it doesn’t fall into the data center accelerator category targeted by US controls since October 2022, according to the observers cited.

An Nvidia spokesperson told Reuters that US companies remain hamstrung by outdated export controls targeting gaming products released nearly five years ago, and by China’s own limits on US imports.

A signal of openness is not a delivery

The H200 precedent calls for caution. In January 2026, Beijing asked Alibaba, Tencent and ByteDance to prepare H200 orders. The United States granted licenses to seven Chinese companies, for up to 75,000 chips each. By May, not a single H200 had been delivered, according to Digital Today. An actual delivery requires three green lights: Beijing’s approval, a US license and no Chinese brake on imports.

The logic behind the drop in Chinese stocks is simple. If the country’s groups can buy Nvidia again, the investment case for local chips, described as “years behind” the United States and Taiwan, weakens. Bloomberg’s Monday headline points to a one-year low for Chinese stocks, with chip and optical names falling.

The stocks mentioned are listed in Shanghai or Hong Kong, and therefore not eligible for the PEA. Nvidia is held in a standard brokerage account.

Sources:

Tesla Optimus: several hundred robots a week according to The Information, but the hands are a bottleneck

Tesla is reportedly now building several hundred Optimus robots a week in Fremont, about ten times more than in the second quarter. These are the first production rate figures to surface, and they come from The Information, not from Tesla.

Key points:

  • According to The Information, production has gone from a few dozen to several hundred robots a week. The goal is more than 1,000 a week by the end of 2026.
  • The hand is the weak spot: the hands and forearms contain more than 100 screws and small parts, assembled by hand.
  • Tesla is considering leasing the robots to a handful of companies rather than selling them, according to Electrek.
  • Tesla shares slipped about 2% on Friday, falling back after opening higher.

What the report reveals

On August 27, Tesla announced that Optimus had entered production without giving a rate. The report fills that gap and describes several hurdles. Assembly jigs drift out of alignment, which increases rework. Tactile sensors aren’t reliable enough: Tesla plans a replaceable “sensing glove” next year, so it no longer has to swap out the entire hand. Suppliers, often Chinese, are struggling to maintain quality at higher volume, especially for motors and precision gears.

On the intelligence side, the robot reportedly takes several days to learn a simple task and behaves unpredictably outside its training scenarios. Hence the leasing strategy, which echoes Full Self-Driving: ship the hardware, collect data, promise better software later. The long-term internal target is reportedly 20,000 robots a week, according to Electrek. Tesla has not confirmed these figures in the sources consulted.

China dominates patents

The competition is also playing out over intellectual property. China held 60.8% of humanoid robot patents filed with the world’s five major patent offices in the first half of 2026, versus 21.8% for the United States and 4.6% for South Korea, according to data from South Korean lawmaker Lee Jeong-heon reported by the Korea Herald. China’s share was 51.8% in 2024. In the first half of 2026, China filed 1,584 applications, compared with 113 for the United States.

These figures measure volume, not quality: the journal IAM finds that China leads on quantity while the United States keeps an edge on quality.

Tesla is held in a standard brokerage account, not in a PEA. Thematic robotics UCITS ETFs are the other way in from Europe.

Sources:

AI chips: Micron reports Wednesday, TSMC reportedly speeding up on 2 nm

Two signals from the chip supply chain this week: Micron’s earnings, expected Wednesday, September 30 after the Wall Street close, and a Taiwanese report raising projections for TSMC’s 2 nm process.

Key points:

  • For its fiscal fourth quarter, Micron guided to revenue of $50 billion (plus or minus $1 billion) and a gross margin of about 86%.
  • The analyst consensus is around $50.4 billion in revenue and $31 in adjusted earnings per share, according to AlphaBetaStock.
  • According to EDN, TSMC could reach about 120,000 2 nm wafers a month by the end of 2026, versus 90,000 to 100,000 previously estimated.
  • These TSMC figures come from the Taiwanese press and are not an official company forecast.

Micron: expectations are already sky-high

Last quarter, Micron posted $41.46 billion in revenue and an 84.6% gross margin (results released June 24). The stock was trading around $974 in mid-September, up roughly 240% since January, according to AlphaBetaStock. HBM memory, which goes alongside AI chips, is sold out for the year under fixed-price contracts, according to AlphaBetaStock. That provides visibility, but it also limits the room for an upside surprise.

Analysts will mainly be watching the gross margin guided for the next fiscal year. One risk cited: Samsung’s requalification at Nvidia, which could end the current HBM duopoly and weigh on prices. At SK hynix, the imbalance between DRAM supply and demand is expected to worsen in 2027 and not ease until 2030 at the earliest, according to the Motley Fool.

TSMC: 2 nm is ramping faster than expected, according to EDN

According to EDN, as reported by Investing.com on September 27, Apple, Nvidia, AMD, Qualcomm and MediaTek recently raised their orders for the 2 nm family by 10% to 20%. Five 2 nm fabs are coming online this year: two in Hsinchu and three in Kaohsiung, according to a TSMC vice president quoted by EDN. First-year output would be 45% higher than 3 nm’s in 2023, with 70% annual growth between 2026 and 2028.

TSMC shares were down about 1% in Taiwan at the time of publication. These developments add to the 5 to 10% price increase for 2027 reported by Nikkei Asia in July, and to the 2026 capital spending budget of $60 billion to $64 billion unveiled with second-quarter results.

Neither Micron nor TSMC is eligible for direct holding in a PEA. Semiconductor UCITS ETFs offer exposure from Europe.

Sources:

Edition prepared by our monitoring agent (data and sources as of Monday, September 28, 2026), reviewed in line with our editorial policy. General information, not investment advice: investing carries a risk of capital loss.