Update from the book

Chips, computing power, energy

Who makes the processors for AI, how much the data centres consume, how export controls are moving.

As of5 October 2026Reviewedevery three monthsIn the book12Italiano

Companies building data centres for artificial intelligence now invest more than the whole world spends on extracting oil and natural gas. That is the scale the chapter is about, said with a single figure. Everything that follows carries its own date, because on this subject a line without a date goes stale in a few weeks.

Who actually ships, and who has only announced

Next to each accelerator is the date of the announcement and today's status. They are two different things: between the presentation on stage and the first rack delivered there are six to eighteen months, and in the meantime the name circulates as if it were already running.

WhoAcceleratorAnnouncedStatus at 3 October 2026
NvidiaBlackwell Ultra, GB30018 Mar 2025shipping in volume, and it is the generation carrying the revenue
NvidiaVera Rubin, with the Groq 3 LPX inference rack31 May 2026ramping into full production, with racks running at CoreWeave, Google Cloud, Microsoft Azure, Oracle and Nebius: Nvidia release of 26 August 2026
GoogleTPU v7 IronwoodApr 2025available on Google Cloud, general availability announced on 6 Nov 2025
GoogleTPU 8t Sunfish and TPU 8i Zebrafish22-24 Apr 2026announced only. Production expected in late 2027
AmazonTrainium32 Dec 2025available, Trn3 servers in general availability
AmazonTrainium4Dec 2025announced only, no availability date
MicrosoftMaia 20026 Jan 2026in production in the Iowa data centres; Arizona declared as next
AMDInstinct MI350 and MI355X12 Jun 2025available from cloud providers since the second half of 2025
AMDInstinct MI455X and Helios rack23 Jul 2026the AMD release contains no availability dates. The first shipments had been stated verbally for the end of September, and at 3 October 2026 no release carrying a date has appeared
CerebrasWSE, the chip the size of a dinner plate-in production. The company listed on Nasdaq on 14 May 2026
MatXaccelerator for language models only-no chip on the market. First shipments expected in 2027
GroqGroqCloud, inference cloud-has moved from chips to selling computing capacity: 13 data centres, including Nvidia systems. It licensed its own technology to Nvidia (see below)

The dash in the date column marks the products the company does not tie to a specific announcement. Cerebras, Groq and MatX are the three startups the book names. Cerebras sells chips, MatX has none on the market yet, Groq has moved to selling computing capacity.

Groq was not bought. On 24 December 2025 Groq and Nvidia announced a non-exclusive licence agreement on inference technology. Founder Jonathan Ross and president Sunny Madra moved to Nvidia with part of the team, but Groq remains an independent company, led by Adam Winter, and GroqCloud keeps running. In June 2026 Groq raised 650 million dollars, and on 17 August another 350 million at a valuation of 3.5 billion, led by Disruptive and with Nvidia expected to take part. In the August release it describes itself as an "NVIDIA Cloud Partner", with 13 data centres and a plan to go from 54 to more than 200 megawatts in 2027, including Nvidia GPU clusters.

What it cost is in Nvidia's accounts. The figure of about 20 billion dollars quoted everywhere is journalistic. In its Form 10-K for the fiscal year ended 25 January 2026, at note 2, Nvidia writes that it paid 13.0 billion at closing and owes another 4 billion within a year, imputed interest included: the recorded cost is 16.9 billion. Of this, 2.5 billion is the licence on the technology, valued at cost to recreate and amortised over five years, and 14.4 billion is goodwill, "primarily attributable to the workforce and future development of the licensed technology". Nvidia states that no customer contracts, existing products or equity interests were purchased. According to the New York Times, as reported by Reuters on 9 September 2026, the US Department of Justice is investigating whether the agreement was structured to avoid antitrust review, and has sent Nvidia a formal request for information. The technical fruit of the agreement is visible: the accelerator born of that licence, the Groq 3 LPX, is part of the Vera Rubin platform Nvidia announced on 31 May 2026. I told the mechanism, with four other similar cases from the last two years, in "Comprare senza comprare" (in Italian), on 4 October 2026.

Market shares, and why none of them is verifiable

The number that always circulates is Nvidia's share of data centre accelerators: you read between 80 and 88%, depending on the research house publishing it. None of those houses publishes its method without a subscription, and Nvidia does not state its own share. TSMC does not publish its own either: the 72-73% you read comes from TrendForce and Counterpoint, which measure different perimeters from each other.

What can be calculated from filed accounts is another thing, and has to be said with its perimeter. Adding up the "data center" segment revenue of the three listed Western suppliers in the latest quarter of each, Nvidia is worth 87.2%: 89.0 billion dollars out of 102.0 in total. The three quarters are not the same calendar quarter: Nvidia closes on 26 July 2026, AMD and Intel on 27 June.

That number measures revenue, not units. Inside the "data center" segment it also includes central processors for servers and networking equipment, not accelerators alone. And above all it excludes the chips the operators build in-house - Google's TPUs, Amazon's Trainium, Microsoft's Maia - because they are not sold and so have no revenue to add up. That is tens of billions of dollars of compute that does not enter the count.

If you see an AI market share without its perimeter next to it, that figure does not tell you what you think.

Where they are made

TSMC makes almost all the latest-generation chips in Taiwan. The breakdown of its revenue by node says how much the most advanced technology really weighs, which in the headlines already looks like the majority.

Process nodeShare of wafer revenue
2 nanometres3%
3 nanometres30%
5 nanometres33%
7 nanometres11%
All the rest23%

TSMC, second quarter 2026 results release, quarter ended 30 June 2026. The 2-nanometre node has been in volume production since the fourth quarter of 2025 and is still worth 3% of revenue. The next one, A16, has volume production expected by the end of 2026 and at 3 October 2026 it does not appear to have started.

Arizona is the case to read with two columns. Between March 2025 and 16 July 2026 TSMC raised its announced investment in the United States to 265 billion dollars, for ten plants plus two advanced packaging facilities and a research centre. Of those ten, today it produces in one, on the 4-nanometre node. The second finished construction in April 2026, is waiting for the machines and should produce at 3 nanometres in the second half of 2027. Intel, for its part, has taken the 18A node into trial production, and there is still no external customer running volumes on it.

Exports: three clocks, and an extension so far only in words

The restrictions the book describes still exist, but almost all of them have been suspended, not revoked, in the trade truce between the United States and China. A suspension has an expiry date, and without an extension the rule comes back into force on its own.

MeasureStatus at 3 October 2026Next date
United States controls on equipment and HBM memory towards China (Oct 2022, Oct 2023, Dec 2024)in force—
Licence review for Nvidia H200 and equivalents towards China and Macaoeased: from presumption of denial to case-by-case review, rule published on 13 January 2026 and effective from the 15th—
United States rule on affiliates 50%-owned by listed entitiessuspended; an extension to 10 January 2027 was announced on 23 September 2026 and at 2 October has not yet been published in the Federal Register9 November 2026
Chinese controls on rare earths and related technologies (Oct 2025)suspended; extension to 10 January 2027 agreed at the September summit, with no new announcement from China's ministry of commerce at 2 October10 November 2026
Chinese export ban towards the United States on gallium, germanium and antimonysuspended; Beijing's licence controls remain27 November 2026

On paper the three suspensions still expire in the same month. If none is extended by a formal act, in November 2026 one American rule and two Chinese ones come back into operation together, and the picture of August 2026 stops holding in three weeks. On 23 September 2026, on the eve of the meeting between Trump and Xi Jinping in Washington, Treasury Secretary Bessent announced that the two countries are extending the trade truce by two months, to 10 January 2027. At 2 October the extension appears neither in the Federal Register nor in a new announcement from China's ministry of commerce, and neither government has said explicitly which of the three suspensions it covers. At the summit, which ended on 25 September, the two presidents also opened a dialogue on AI and set a new meeting for November.

A regulatory opening is not yet a market: at the end of February 2026 Nvidia obtained a licence for a limited volume of H200 towards China, and in the release of 26 August 2026 it repeated that it was counting no revenue from data centre compute in China in its own outlook. On the minerals side, on 2 February 2026 the United States export-import bank approved up to 10 billion dollars for Project Vault, the strategic reserve of critical minerals, which with about 2 billion of private capital reaches the 12 billion stated.

Energy, and what the bill costs

The International Energy Agency revised its own estimates on 16 April 2026, with a report that updates the one from April 2025 from which almost all the figures in circulation come.

World electricity consumption of data centresValue
Consumption in 2025485 TWh
Growth in 2025+17%
Projection to 2030about 950 TWh
Share of world electricity demand in 2030about 3%
Projection to 2035about 1,200 TWh
Emissions in 2035about 350 million tonnes

The most useful figure in the new report concerns AI alone: data centres dedicated to artificial intelligence consumed 50% more in 2025 and their demand triples between 2025 and 2030, much faster than the total. It is a measure the agency signs, unlike the ranges on AI's share inside data centres, which come from commissioned work and swing between 5 and 20%.

On spending, the same agency writes that the capital invested by the five largest technology groups passed 400 billion dollars in 2025, that in 2026 it is expected to grow by another 75%, and that today it exceeds world investment in extracting oil and natural gas. The four companies that publish a forecast for 2026 put it like this:

Investment announced for 2026Billions of dollarsWhere it is stated
Alphabet195-205verbally, on the analyst call
Amazonabout 200verbally, on the analyst call
Microsoftabout 175verbally, on the analyst call
Meta130-145in the filed release

The total is around 700 billion dollars, but it is a sum of midpoints and should be taken as an order of magnitude. Only one of the four figures, Meta's, appears in a filed document: the other three are spoken and picked up by the financial press. Microsoft's figure, moreover, is not comparable with the one from a few months earlier, because the company extended the estimated useful life of its buildings from 15 to 25 years and moved part of its lease contracts out of the investment count.

Water, and what that estimate really measures

The figure you find everywhere is between 312 and 765 billion litres of water consumed by AI systems in 2025. It comes from a study by Alex de Vries-Gao published in Patterns on 17 December 2025, and it is a serious estimate. It is worth knowing what is inside it.

It is a modelled projection, not a measurement: it starts from an estimate of the electrical power drawn by AI, crosses it with the environmental reports of Apple, Google and Meta and with coefficients from the literature on water and emissions per kilowatt-hour. It covers both the cooling water of the data centres and the water consumed by the power stations that produce the electricity, and in the author's count the water of the power stations is the dominant item.

The reason the range is a factor of two and a half wide is written in the study: operators do not separate AI workloads from the others in their own environmental reports, and the share attributed to AI is approximated from general data centre metrics. Google, in a note to a report on the impact of Gemini, wrote that it did not want to report indirect water because it does not control the power stations that consume it on its behalf. As long as the data are published only by the companies, and in this form, a single number on this subject does not exist.

The bottleneck, where it is now

The book's criterion asks you to look at which resource is missing. At 3 October 2026 three are missing at once, and none of the three is a chip.

The electricity grid. At the end of 2025 there were about 8,200 projects queuing for connection in the United States, for 1,312 gigawatts of generation and another 749 gigawatts of storage. Most of the projects that enter the queue are then withdrawn, and those that reach the end take on average longer than before. Hence the data centre projects with their own on-site gas plant, which the energy agency records as a direct consequence of the slowness of connections: according to the satellite tracking the agency cites, about a fifth of those projects have already started land clearing or construction. Global Energy Monitor measures it: in August 2026 the gas capacity tied to United States data centres, across announced, pre-construction and construction, is 189 gigawatts, nearly double what it was six months earlier, inside a total of planned gas plant that has gone from 252 to 378 gigawatts. Of the 52 gigawatts actually under construction, 16.9 are for data centres; Texas alone is developing 122, of which 77 for data centres.

Memory. Over the past six months a shortage of high-bandwidth memory has opened up, the HBM that sits inside every accelerator, and the energy agency does not expect it to be resolved before the end of 2027. It is the constraint that today limits how many accelerators can be built, upstream of foundry capacity.

The consent of local communities. In the summer of 2026 several American local administrations voted on data centres, and the picture is not one-sided: on 16 June the city council of San Marcos, Texas, removed data centres from the zoning plan, making them inadmissible across the whole municipality; on 8 July the planning commission of Fort Worth rejected its own regulation and sent it back to the council; on 14 July the commissioners of Stokes County, North Carolina, instead rejected a six-month moratorium. Anyone telling this story as a single front of opposition is simplifying: the votes go in both directions, and it is the fact that there is a vote at all that is new. Between 14 August and 3 October the votes again went in both directions. On 8 September the board of Lake County, Illinois, unanimously approved an eight-month moratorium on data centres in unincorporated areas, until 11 May 2027; on 30 September the city council of Palm Springs, California, unanimously extended its own moratorium by twenty-two and a half months. In the opposite direction, in mid-September the city commission of Clovis, New Mexico, rejected a 120-day moratorium by 5 to 3, and on 22 September the commission of Raton, in the same state, rejected a one-year moratorium by 3 to 2, despite a petition with 477 signatures. In the revision of 19 September I wrote that I had found no vote in the opposite direction: the Clovis vote had already taken place.

On nuclear, the most quoted project is the restart of Three Mile Island Unit 1, renamed the Crane Clean Energy Center, for 835 megawatts destined for Microsoft's data centres. Constellation states it will come online in 2027, and the nuclear safety authority's decision on the operating licence is expected in May 2027. At 3 October 2026 there is not a kilowatt-hour on the grid.

What an hour of GPU costs

The book gives no compute prices, and the reason is visible as soon as you put two public price lists side by side. These are the pay-as-you-go prices of two specialist providers, read on the same day.

AcceleratorProviderDollars per GPU per hour
H100Lambda3.99
H100CoreWeave6.155
H200CoreWeave6.305
B200Lambda6.69
B200CoreWeave8.60
GB300CoreWeavenot published

Public pay-as-you-go price lists, re-read on 3 October 2026 and identical to those of 14 August, in dollars because neither provider publishes in euros. At the European Central Bank reference rate of 2 October 2026, one euro is worth 1.1225 dollars: the same H100 therefore costs about €3.55 an hour from one provider and €5.48 from the other. Prices with advance reservation fall by more than half.

The very same accelerator costs over 50% more moving from one price list to the other, and the most recent model has no public price at all. If you need an order of magnitude for a quote, take the range; if you need a number, ask the provider with the exact configuration and the date.

What has changed

5 October 2026 · revision

The figure for the Nvidia-Groq agreement is in Nvidia's accounts, and it was missing here. Until 3 October this page said the figure appears in no release. For the releases that was true, but Nvidia's annual report gives it: 16.9 billion dollars, of which 14.4 billion is goodwill. I have added Groq's 650 million round of June, the Department of Justice investigation reported by the press in September and a link to the article "Comprare senza comprare" (in Italian). Groq is also back in the startup table, from which it had dropped out: it now sells computing capacity.

3 October 2026 · revision

The extension of the three suspensions has been announced and is waiting for the written act. On 23 September 2026 Treasury Secretary Bessent announced that the United States and China are extending the trade truce by two months, to 10 January 2027, and the summit between Trump and Xi ended on 25 September. At 2 October the new date appears neither in the Federal Register nor in an announcement from China's ministry of commerce, and neither government has said which of the three suspensions it covers: in the table the expiry dates remain those in force, with the announced extension next to them.

Local votes are again going in both directions. On 30 September Palm Springs, California, extended its own moratorium on data centres; in September Clovis and Raton, New Mexico, each rejected one. In the revision of 19 September I wrote that I had found no vote in the opposite direction to the Lake County moratorium: the Clovis vote, in mid-September, had already taken place.

Correction on on-site gas. This page attributed to about a fifth of new United States data centres the start of work on their own gas plant. In the energy agency's report that fifth refers to the data centre projects with on-site gas generation that have already started land clearing or construction.

Correction on Vera Rubin. Nvidia's release of 26 August 2026 says "ramping into full production", and the table and the 19 September entry now say the same.

A Groq round that was missing. On 17 August 2026 Groq raised 350 million dollars at a valuation of 3.5 billion, with Nvidia expected to take part. It predates the revision of 19 September and was not on the page.

What I did not update, and why. TSMC reports the third quarter on 15 October, so the node table stays at the second. Nvidia reports in November, AMD and Intel at the end of October: the share calculated on filed accounts stays at 87.2%. The International Energy Agency has published nothing new on AI since 19 September. The CoreWeave and Lambda price lists, re-read on 3 October, carry the same prices; only the exchange rate changes, now taken at 2 October. AMD has not yet published a release with shipping dates for MI455X and Helios. None of the four companies has revised its 2026 investment guidance.

19 September 2026 · revision

Vera Rubin is in production, and Nvidia's share stays where it was. On 26 August 2026 Nvidia reported 89.0 billion dollars of data centre segment revenue, out of 96.2 billion in the quarter ended 26 July, and stated that the Vera Rubin platform is ramping into full production, with racks running at five cloud providers. Redoing the calculation on the latest filed quarter of each of the three, Nvidia's share goes from 87.3 to 87.2%: a tenth of a point, across three entirely new quarters.

The three November clocks run through 24 September. Xi Jinping is expected in Washington from 23 to 25 September 2026, with the meeting with Trump on the 24th. At 19 September none of the three suspensions - the American affiliates rule, the Chinese controls on rare earths, the ban on gallium, germanium and antimony - appears to have been extended, and no new act appears either in the Federal Register or among Beijing's communications.

Gas is growing faster than grid connections, and now it has a number. Global Energy Monitor measures 189 gigawatts of gas capacity tied to United States data centres in August 2026, nearly double what it was six months earlier. It is the same thing the energy agency described in words, with a source that can be opened.

Correction. This page had Groq led by Simon Edwards. Groq's own people page, consulted on 19 September 2026, states Adam Winter as chief executive, in post since 2026.

What I did not update, and why. The International Energy Agency published nothing new on AI between 14 August and 19 September: the energy figures remain those of 16 April 2026. TSMC has not yet reported the third quarter, so the node table stays at the second. The CoreWeave and Lambda price lists, re-read today, carry the same prices as in August. None of the four companies has revised its 2026 investment guidance. On water, no new study has appeared.

14 August 2026 · first publication

The Chinese restrictions on minerals are paused, not lifted. The book describes them as in force, and that was true until November 2025. Since then they have been suspended, with expiry dates of 10 and 27 November 2026. Anyone designing a supply chain that passes through those materials has a horizon of a few weeks, not of years.

The energy projection has been rewritten from a higher base. The energy agency has moved from 415 terawatt-hours in 2024 to 485 in 2025 as its starting point, and the estimate for 2030 is about 950. The final total moves little; what moves is the speed at which it is reached, and the share of that total that belongs to AI alone, which triples.

Two of the three chip startups named in the book have changed nature. Cerebras listed on Nasdaq on 14 May 2026, after a placement at 185 dollars a share. Groq has licensed its own technology to Nvidia on a non-exclusive basis and has lost its founders, while remaining independent. The third, MatX, raised 500 million in February 2026 and still has no chip on the market.

The energy agency's report →

Three numbers you will not find here, and why.

A market share from a primary source does not exist: no manufacturer publishes it and whoever calculates it sells the data. The most that can be verified is the calculation on filed accounts reported above, with the perimeter next to it.

The stock market capitalisation of Nvidia changes every day and no company publishes it: if you need it, look it up on a stock exchange listing with the time next to it, not on a page like this one.

The hourly price lists of the three big cloud providers exist but can only be read through their interactive calculators. I have not reported them because I cannot cite a dated version, and a price without a date on this subject lasts a few weeks.

Sources

  • Nvidia, first quarter fiscal 2027 results release, 20 May 2026; "Vera Rubin" release of 31 May 2026; second quarter fiscal 2027 results release, 26 August 2026.
  • Groq, release of 24 December 2025 on the non-exclusive licence agreement with Nvidia. The figure of about 20 billion dollars is journalistic in origin (CNBC, 24 December 2025). The recorded cost, 16.9 billion, is in Nvidia's Form 10-K for the fiscal year ended 25 January 2026, note 2, read on 5 October 2026.
  • Groq, releases of 22 June 2026 on the 650 million dollar round and of 17 August 2026 on the 350 million Series A round, consulted on 5 October 2026.
  • Reuters, 9 September 2026, on the Department of Justice investigation reported by the New York Times; Nvidia, product page for the Groq 3 LPX. For the account of the deal: Francesco Masera, "Comprare senza comprare", Libro Vivo IA, 4 October 2026 (in Italian).
  • Release announcements from Google Cloud (6 November 2025), Amazon (2 December 2025), Microsoft (26 January 2026) and AMD (12 June 2025 and 23 July 2026).
  • Cerebras Systems, release on the pricing of the initial public offering, 13 May 2026.
  • TSMC, second quarter 2026 results release, and SEC filing of 3 March 2025 on the Arizona investment. The expansion to 265 billion was announced on the analyst call of 16 July 2026, without a dedicated release. Newsroom and quarterly results page re-read on 3 October 2026.
  • Bureau of Industry and Security, "Revision to License Review Policy for Advanced Computing Commodities", Federal Register, 15 January 2026; BIS release of 2 December 2024.
  • Export-Import Bank of the United States, fact sheet on Project Vault, 2 February 2026.
  • International Energy Agency, "Key Questions on Energy and AI", 16 April 2026, summary. Updates "Energy and AI" of 10 April 2025, which remains published. Summary re-read on 3 October 2026.
  • Lawrence Berkeley National Laboratory, "Queued Up: 2026 Edition", data at the end of 2025, published in June 2026, page re-read on 3 October 2026.
  • de Vries-Gao, A., Patterns, 17 December 2025, DOI 10.1016/j.patter.2025.101430, and release from the Vrije Universiteit Amsterdam.
  • Constellation Energy, page of the Crane Clean Energy Center, consulted on 14 August 2026; newsroom re-read on 3 October 2026.
  • Public price lists of CoreWeave and Lambda, re-read on 3 October 2026; European Central Bank, reference rate of 2 October 2026.
  • AMD, second quarter 2026 results release, 4 August 2026; Intel, second quarter 2026 results release, 23 July 2026. These are the two quarters that enter the share calculation alongside Nvidia's.
  • Global Energy Monitor, "U.S. gas power proposals tied to data centers nearly double in six months", August 2026, re-read on 3 October 2026.
  • Daily Herald, 9 September 2026, on the Lake County board vote of 8 September (the county's own bulletin reports the recommendation of the Zoning Board of Appeals).
  • KESQ, 30 September 2026, on Palm Springs; Eastern New Mexico News, 23 September 2026, and ABC7 Amarillo, 19 September 2026, on Clovis; KRTN, 22 September 2026, on Raton. Consulted on 3 October 2026.
  • Groq, people page of the company website, consulted on 19 September 2026.
  • On the Trump-Xi summit and the extension of the truce: China Briefing, 29 September 2026; Brookings Institution, "What did the Trump-Xi summit actually accomplish?"; The Nation Thailand; National Law Review, "BIS Affiliates Rule Suspension Extended but Uncertainty Remains", 2 October 2026. Consulted on 3 October 2026. At 2 October neither administration has published a formal act with the new date.
  • On the Chinese suspensions I did not reach the original text of Beijing's ministry of commerce: the dates come from several independent reconstructions by international law firms, consistent with one another.

Where you come from

  • Who owns artificial intelligence

Update history

  1. 5 October 2026Groq box corrected: the figure for the agreement with Nvidia does not appear in the releases but is in Nvidia's Form 10-K, note 2: a cost of 16.9 billion dollars, of which 14.4 billion goodwill and 2.5 billion licence. Added Groq's 650 million round of June 2026, the Department of Justice investigation reported by Reuters on 9 September 2026 and a link to the article "Comprare senza comprare" of 4 October 2026. Groq is back in the startup table as an inference cloud; the accelerator in the Vera Rubin row takes its exact name, Groq 3 LPX.
  2. 3 October 2026Revision of 3 October 2026, window from 19 September, mirroring the Italian page paragraph by paragraph. On 23 September the United States and China announced the extension of the trade truce to 10 January 2027, which at 2 October has not yet appeared in a formal act: in the table the expiry dates remain those in force, with the extension next to them. Added the local votes of Palm Springs on 30 September, Clovis in mid-September and Raton on 22 September, and corrected the 19 September sentence that found no vote against the Lake County moratorium. Corrected the reading of the energy agency on on-site gas: the fifth refers to projects with their own plant that have already started work. Corrected the verb on Vera Rubin, ramping into full production according to Nvidia's release of 26 August 2026. Added Groq's round of 17 August 2026, 350 million dollars, which was missing. GPU price lists re-read and identical, ECB rate moved to 2 October; the bottleneck and Three Mile Island lines, left at 14 August, have been re-checked and now carry the date of 3 October.
  3. 27 September 2026Book references aligned to the printed fourth edition (thirteen chapters): the chapter that points here is 12, Who owns artificial intelligence, not 13. Data and text unchanged.
  4. 19 September 2026Revision of 19 September 2026, window from 14 August, mirroring the Italian page paragraph by paragraph. Vera Rubin is in full production, with racks running at five cloud providers: Nvidia release of 26 August 2026, which also takes data centre segment revenue to 89.0 billion dollars and redoes the share calculated on filed accounts, from 87.3 to 87.2% across three entirely new quarters. The three export suspensions are still standing and run through the Trump-Xi meeting of 24 September. Added Global Energy Monitor's measurement of gas capacity tied to data centres, 189 gigawatts in August 2026. Corrected the name of Groq's chief executive, from Simon Edwards to Adam Winter, from the company's own people page. GPU price lists re-read and identical to August, ECB rate moved to 18 September. Also fixed a pre-existing typo in the GPU pricing section ("50% more more moving"). Not updated because nothing new appeared: IEA energy figures, TSMC nodes, water, 2026 investment guidance. Stated: the two lines still carrying the date of 14 August - the bottleneck and the Three Mile Island kilowatt-hour - were not re-measured against a primary source in this pass.
  5. 8 September 2026First English version, translated from the Italian page as of 14 August 2026.

This page replaces a list that would have aged on paper. The date says when I last reviewed it; what changed is at the bottom.

Printed from https://librovivoia.it/en/r/hardware · As of 5 October 2026