The chipmaker finances companies that need its hardware, takes
stakes in the businesses building demand for it and has even agreed to
buy cloud capacity those businesses can’t sell. Meanwhile, people trying
to buy graphics cards around the world are learning how little choice a
market can offer when one supplier holds most of it.
15 November 2026
When the price of a graphics card goes up, somebody will tell you
it’s supply and demand. Fine. Who is financing the demand? Who has the
money to reserve manufacturing capacity years ahead? And who gets told
to wait until the price comes down?
Marks & Spencer could write a cheque to an AI start-up and hope
to make money on its stake. When Nvidia writes that cheque, it can also
make money selling the equipment the start-up needs to spend it. The
company has every incentive to make that equipment the foundation for
the customer’s next expansion. That is a far more consequential use of
corporate capital than buying a share in a promising business. It is a
supplier helping build its own market.
Follow the money
| Deal | What the public record shows |
|---|---|
| OpenAI, February 2026 | OpenAI announced $30 billion in Nvidia investment alongside an agreement for next-generation Nvidia computing capacity. OpenAI announcement |
| Anthropic, November 2025 | Nvidia announced a commitment of up to $10 billion. Anthropic committed to an initial deployment of up to one gigawatt of compute on Nvidia Grace Blackwell and Vera Rubin systems through Microsoft Azure. Joint announcement |
| SB Energy and OpenAI, August 2026 | Nvidia disclosed credit guarantees capped at $105 billion for land, power and construction at an Ohio campus covering approximately 4.25 gigawatts of IT load. Under 20-year leases to OpenAI, the campus is to run Nvidia infrastructure exclusively, subject to limited exceptions. The guarantees take effect in phases as conditions are met. Nvidia SEC filing |
| CoreWeave, September 2025 and January 2026 | Nvidia agreed to purchase eligible residual unsold cloud capacity through April 2032 under an order with an initial value of $6.3 billion, subject to contract conditions. Four months later it announced a $2 billion investment in CoreWeave shares. CoreWeave 8-K, capacity agreement · SEC filing, equity investment |
CoreWeave is the clearest example of how many sides of a single
market Nvidia can occupy. It supplies the hardware platform, owns shares
in the cloud operator and has taken on an obligation to buy eligible
capacity that other customers leave unsold. CoreWeave told Reuters that
the equity investment would be used for expansion costs such as land and
power, rather than buying Nvidia processors with that cash. A powered
data centre is precisely what allows an Nvidia-based cloud operator to
deploy more Nvidia hardware. Meanwhile the capacity agreement gives
CoreWeave a buyer of last resort for capacity covered by the contract.
The $6.3 billion was the initial order value, with actual obligations
subject to the agreement’s terms. CoreWeave
8-K · Reuters
The Ohio guarantee puts the scale of the strategy in plain sight. The
$105 billion is a ceiling on conditional credit support, rather than a
sum already paid out. The terms link the financing of a physical site to
long leases for a customer and infrastructure from the company providing
the guarantee. Ask what an independent chip supplier has to do to
compete at a site whose financing and equipment are already linked in
this way. Then ask which companies could afford to offer those terms in
the first place. Nvidia
SEC filing · Reuters video
The scale goes beyond any single deal. Nvidia’s July 2026 quarterly
filing recorded $31 billion in net additions to its
non-marketable equity securities over six months, compared with
$948 million in the same period a year earlier. These are principally
private-company holdings. The figure covers the portfolio and gives no
deal-by-deal accounting of how much ultimately went back into Nvidia
products. It does show how quickly taking stakes in other businesses has
grown alongside selling them hardware. Nvidia
10-Q, note 6
There’s a familiar defence for these deals. Large projects need
financing, guaranteed customers make infrastructure easier to build and
cheaper compute can benefit the people using it. Fine. The question is
what that financing buys the supplier. A rival entering this market has
to compete on hardware, software, networking and its ability to support
a deployment. It may also have to compete with a supplier willing to put
billions into the customer’s business or assume some of the risk of
unused capacity. A technically excellent chip can still lose that
contest.
What gamers are paying
The price of a graphics card can rise even if it costs exactly the
same to make. If suppliers send fewer cards into the retail channel
while people still want them, sellers can charge more. Nvidia can
collect more if wholesale prices rise. Board partners, distributors and
retailers can collect more at their own stages. Each can point to the
final price as a market outcome. The only person who can’t pass it on is
the buyer at the till.
There’s a documented link between the wider AI buildout and consumer
supply. Reuters reported in February that AI demand was pushing up
memory prices and prompting manufacturers to favour components for
higher-margin data-centre chips. Nvidia itself warned of supply
constraints in gaming. AI accelerators and gaming cards use different
kinds of memory and packaging, so a data-centre GPU does not translate
into one fewer GeForce card. The pressure comes where manufacturers can
allocate investment or capacity towards the more profitable products.
Higher component costs and higher markups can also occur at the same
time. One should never be assumed from the other. Reuters
· Nvidia
annual report
TechSpot tracked the cheapest in-stock gaming graphics cards across
ten markets: the US, Canada, Australia, Germany, Britain, India, the
Philippines, Brazil, Poland and the Netherlands. Its survey put the
average rise at 15% in August, 7% in September and 11% in
October, or about 35% over three months. By
October, the cards surveyed were averaging 44% above their
estimated local launch prices. That is the average across the
markets sampled. Individual cards and countries varied considerably. TechSpot
international pricing survey
The British figures show how that played out at the till. TechSpot
recorded the cheapest card it could find in stock for each model in its
September and October surveys. These are snapshots of the price a
shopper could actually find, with the usual limitations of retailer
coverage and changing stock.
| Card | September UK snapshot | October UK snapshot | Increase |
|---|---|---|---|
| GeForce RTX 5090 | £4,300 | £5,600 | £1,300 |
| GeForce RTX 5070 Ti | £1,000 | £1,130 | £130 |
| GeForce RTX 5070 | £700 | £740 | £40 |
| Radeon RX 9070 XT | £705 | £760 | £55 |
| Radeon RX 9060 XT 16GB | £460 | £510 | £50 |
The top-end card makes the most outrageous headline, but the £50
increase on a mid-range Radeon is arguably the more relevant figure for
most shoppers. It shows the squeeze crossing brands. PC Prowler’s UK
index, which tracks prices on comparable in-stock listings, reached
125.9 in November to date, against 100 in
March. These are asking prices and indices. To calculate who
captured the extra money, you would need wholesale invoices, channel
margins and volumes. To calculate how much Nvidia’s own investments
added to the price, you would also need to know how much computing
demand and production allocation would have changed without those deals.
PC Prowler UK GPU
index
Nvidia’s financial results explain where the incentives lie. It made
$89 billion of its $96.2 billion quarterly revenue from data
centres in the quarter to 26 July. Its company-wide gross
margin was 75%, and Nvidia attributed the year-on-year
increase to a more profitable product mix including Blackwell Ultra.
That’s a company-wide measure of gross profit as a share of revenue. It
can’t tell us the markup on an individual chip. It can tell you which
business is setting the direction of the company. Nvidia
quarterly results · Nvidia
10-Q
The memory makers
are making the same choice
You can refuse to buy a GeForce. Try building a PC without RAM or
somewhere to store your files. That’s where the argument stops being
about one company. The AI boom is also changing who gets served first by
the companies that make the components every PC needs.
Samsung, SK hynix and Micron accounted for 87.6% of global
DRAM revenue in the second quarter of 2026, according to
TrendForce: 39.4%, 24.9% and 23.3%, respectively. That covers the DRAM
market, including server and high-bandwidth memory. The label on a kit
of desktop RAM can change while the chips inside still come from one of
a handful of suppliers. TrendForce
DRAM market analysis · Yonhap
market-share breakdown
Micron gave the clearest account of who comes first.
In December 2025 it announced that it would withdraw its Crucial-branded
RAM and SSDs from consumer channels, with shipments continuing through
February 2026. Its stated reason was to improve supply and support for
larger customers in faster-growing segments, citing AI-driven demand for
memory and storage. Micron still sells chips to partners who put them
into consumer products. What disappeared was a major manufacturer’s own
retail line, because the company saw a better opportunity serving bigger
buyers. Micron
announcement · Micron annual
filing
SK hynix reported an 80% operating
margin for the third quarter of 2026 and said long-term
agreements covered more than half of its 2027–28 production
capacity. It also reported expanding sales of HBM4 and
high-value server products. Those are company-wide figures and
commitments. They don’t tell us the markup on a desktop memory stick.
They do tell us which part of the market the company is preparing to
serve, and how much future capacity is already covered by long-term
deals. TrendForce reported that suppliers were prioritising advanced
capacity for high-performance server products even as the broader DRAM
market remained undersupplied. SK hynix
results · TrendForce
fourth-quarter outlook
NVMe drives are part
of the same squeeze
An NVMe SSD stores data in NAND flash. Its controller and interface
also matter, but the NAND is the component at the centre of this pricing
story. Counterpoint estimated that Samsung, SK hynix, Micron,
Kioxia and SanDisk accounted for 87% of global NAND revenue in
the second quarter. China’s YMTC accounted for another 14%, with
rounding explaining the total. Buying a different SSD brand can get you
a better deal at the till. It may still put money in the hands of the
same NAND producer. Counterpoint
global NAND shares
The companies are benefiting in ways we can name. TrendForce reported
that Samsung gained from higher NAND selling prices and
a richer mix of enterprise SSDs. SK hynix, through its
group including Solidigm, benefited from strong orders for high-capacity
enterprise drives. Micron, Kioxia and
SanDisk all posted sharply higher NAND revenue in the
second quarter as selling prices rose, with output and shipment changes
also affecting their results. By November, TrendForce reported that
AI-backed enterprise SSDs were still leading contract price increases
while consumer demand was weak. When consumers are pulling back and the
better-paying enterprise market is still bidding for supply, waiting for
the next sale only gets you so far. TrendForce
NAND supplier results · November
NAND market bulletin
At retail, MemRadar tracked the same products on Amazon US from
October 2025 to October 2026. The median same-product rise was
323% across 61 DDR5 kits, 107% across 27 DDR4
kits and 144% across 68 NVMe drives. The
median price of the same 24 tracked 2TB drives went from about
$140 to $360. Britain saw severe pressure too. Forge PC
reported a mid-November Amazon UK DDR5 tracker reading of £18.48
per gigabyte, against £6.16 a year earlier.
These are local retail samples. The worldwide evidence comes from the
producers’ market shares and industry contract data, which measure
different things from prices at checkout. MemRadar October
record, using Keepa data · Forge
PC UK market report
This industry has
a record worth remembering
In 2005, Samsung agreed to plead guilty and pay a 300millioncriminalfine * *intheUSoveraninternationalconspiracytofixDRAMprices.Hynixpleadedguiltyandwassentencedtopay * *185
million. The case concerned conduct from an earlier era. For
2026, the evidence here supports scrutiny of supply allocation, customer
commitments and who earns the margin. A fresh cartel allegation would
require evidence of current coordination. US
Department of Justice
Competition authorities should demand the figures that the public
can’t see: allocation by customer segment, the duration and terms of
supply contracts, real volumes going into the PC market and the margins
earned between chipmaker and retailer. Where the evidence warrants
enforcement, use it. Where governments subsidise new capacity, they
should attach measurable conditions that improve access for smaller
buyers and publish who gets served. Otherwise public money can help
build the next factory while the largest customers have already booked
the output.
AMD is competing for the
same AI money
AMD’s response to Nvidia has been to give two of the biggest
potential AI buyers a financial reason to order its chips. Its OpenAI
agreement covers deployment of six gigawatts of AMD Instinct GPUs and
includes a warrant to buy up to 160 million AMD shares at one
cent each. Meta received its own warrant on the same headline
terms. Meta made a binding commitment for an initial gigawatt, with full
vesting contingent on purchases equivalent to six gigawatts. Both
warrants are subject to purchase and share-price milestones and other
conditions. The penny is the exercise price under the warrant. The
economic incentive comes from the potential value of those shares if the
conditions are met. AMD
annual filing, OpenAI · AMD
8-K, Meta · AMD
quarterly filing, both warrants
That gives big AI buyers another place to go, which can put pressure
on Nvidia. It also shows what these buyers can command. A corporation
ordering chips by the gigawatt can get an opportunity to own part of its
supplier. A household buying one card waits for a sale.
AMD’s November results show the change in priorities across the
industry. Quarterly data-centre revenue reached $8.5 billion, up
97% year on year, while gaming revenue fell 40% to $773
million. AMD says the gaming decline reflected lower
semi-custom and graphics sales. Gaming includes console-related business
as well as graphics, so those figures give a view of the company’s
incentives and revenue mix, with limited visibility into Radeon
shipments alone. AMD
third-quarter results
Intel still offers Arc cards. It also sold $5 billion in
Intel shares to Nvidia in December 2025 after the companies
announced plans for data-centre processors and PC chips incorporating
Nvidia RTX graphics chiplets. Intel can sell buyers a genuine
alternative in a discrete Arc card while partnering with Nvidia
elsewhere. The relationship is another reason to judge the card and the
deal in front of you, rather than treating a badge as a vote for an
entirely independent industry. Intel
partnership announcement · Intel
8-K, completed share sale
Where is the competition?
Jon Peddie Research estimated 12.5 million PC graphics add-in boards
shipped in the second quarter of 2026. Reporting on its figures put
Nvidia’s share at about 90%. That measure covers
shipments of discrete desktop graphics boards. It gives a sense of how
much room AMD and Intel have to challenge Nvidia in the market that
matters to somebody upgrading a desktop. Jon
Peddie Research · Tom’s
Hardware
The first chart counts desktop graphics-board shipments. The second
uses DRAM producer revenue. They describe two different markets using
different measures. PC
Gamer reporting on Jon Peddie Research · TrendForce
and Yonhap
Nvidia says regulators in the EU, US, UK, China and South Korea have
requested information on matters including GPU sales, supply allocation,
and its investments and agreements with AI model developers. Those
requests show that regulators are asking questions about the same
conduct. The public documents cited here establish deals between
suppliers and customers. A price-fixing allegation against rival GPU
suppliers would require evidence of agreement or coordination. Rising
prices across rival brands can follow shared supply pressure and common
commercial incentives. Nvidia
10-Q, risk factors · CMA
guidance on horizontal agreements
The questions regulators should ask are specific. Do investment terms
penalise customers for switching to another supplier? Can an AI company
freely move workloads to a rival platform? What share of demand is
supported by equity, warrants or capacity guarantees from the suppliers
themselves? And as those suppliers reserve production for AI, what
happens to the volume and wholesale price of gaming cards? Answers would
tell us whether the market is rewarding better products, giving
entrenched firms another way to keep rivals out, or doing both.
The failure of governments
The failure of ideology is believing that a market stays competitive
by default. The failure of government is seeing how much power has
accumulated, using public money to expand the industry anyway, and
leaving the public to guess who will get the benefit. The US Commerce
Department awarded Micron up to $6.165 billion in
direct CHIPS incentives for projects in Idaho and New York. The stated
goal was to strengthen domestic memory manufacturing. That makes it
reasonable to ask what taxpayers will get beyond extra production on US
soil. Who gets access to it, on what terms, and can the public see if
the promised benefits materialise? US
Commerce Department award
Regulators have already identified ways this problem can spread. The
US Federal Trade Commission’s 2025 staff report examined Microsoft’s
relationship with OpenAI and Amazon’s and Google’s relationships with
Anthropic. It identified potential effects on access to compute and
talent, switching costs and access to sensitive information. In Britain,
the CMA’s public cloud investigation recommended prioritising
consideration of special regulatory status for Amazon Web Services and
Microsoft. Cloud services matter here because once ownership of the
hardware is out of reach, these are among the companies people and
businesses may have to rent from. FTC
staff report · CMA
final cloud market investigation
Governments should publish the obligations attached to chip subsidies
and report whether recipients meet them. They should fund affordable
computing access for schools, public researchers and small businesses.
Competition authorities should test whether financing arrangements,
exclusive deployments or long-term supply contracts leave rivals a fair
chance to compete. They should publish what they find and enforce the
law where the evidence warrants it. If each government treats the public
interest as somebody else’s job while it competes for the next factory
or data centre, the same companies get to set the terms in country after
country.
Watch the
reporting and the companies’ responses
Reuters on Nvidia’s OpenAI data-centre guarantee
How Nvidia provides credit support for a site slated to run its own infrastructure. Check the cap and conditions against the SEC filing.
Bloomberg on Nvidia’s $2 billion CoreWeave investment
How the equity investment sits alongside the cloud relationship and earlier capacity agreement.
CNBC on third-party financing for AI infrastructure
What bringing major asset managers into the buildout could mean for Nvidia’s position. This is analysis of third-party financing, distinct from Nvidia’s own equity investments and guarantees.
YTN on Nvidia’s response to circular-financing criticism
YTN reports on Nvidia’s defence of its AI investments. Yahoo Finance carries Colette Kress’s own words from the earnings call.
Gamers Nexus on Micron and the memory industry
A consumer-hardware perspective on the Crucial exit. Use Micron’s statement and the historical DOJ records for the underlying evidence.
Where to write
Report a market-wide concern to the authority where
you live or do business. Write separately to your elected representative
if you want a policy change, public hearings or disclosure on a subsidy.
Use the company’s published customer service or investor relations
channels to ask about a specific decision. A retailer can answer for a
sale or an advertised price. A chipmaker can answer for its decisions on
production and distribution.
| Jurisdiction | Competition authority and public route |
|---|---|
| United Kingdom | Report a market concern to the CMA · Find and write to your MP |
| European Union | European Commission competition complaints and market information · European Parliament petitions route |
| United States | Department of Justice Antitrust Division reporting · Find your US Representative |
| Canada | Competition Bureau complaint form |
| Australia | ACCC guidance and reporting route for anti-competitive behaviour |
| Brazil | CADE’s Clique Denúncia reporting portal |
| India | Competition Commission of India guide to filing information |
| Japan | Japan Fair Trade Commission procedures for notifications and investigations |
What to include. Identify the exact product or
business practice, your jurisdiction and the dates involved. For prices,
include model numbers, dated screenshots or invoices, and whether the
listed product was in stock. For supply or financing concerns, cite the
underlying filing and the specific clause or disclosure. If you run a
small business and a distributor changed your allocation or terms,
describe what changed and keep the correspondence. Each piece of
evidence should make clear what you observed and what you want the
authority to investigate.
A short note you can adapt
I am asking your office to examine whether hardware supply
agreements, supplier financing and capacity reservations are restricting
access to competitively priced GPUs, RAM and SSDs for households and
small businesses. The attached material includes dated local price
observations and published company disclosures. Please explain whether
your office is examining these arrangements, what information you can
require from suppliers, and whether you will publish findings. Where
taxpayer funding is involved, please identify the access conditions
attached to it and explain how compliance is measured.
Should you buy a different
card?
Yes, put Radeon on the shortlist. For ordinary PC
gaming, it can offer better value at actual UK prices. TechSpot’s
52-game comparison put the RX 9070 XT 24% ahead of the RTX 5070
at 1440p on average. At the roughly £820 and £780 prices in
Forge PC’s 13 November snapshot, the Radeon cost about 5% more and
delivered about 15% lower cost per frame using that benchmark average.
Check the games you play. In TechSpot’s Indiana Jones test with full ray
tracing, the RTX 5070 pulled ahead. TechSpot
benchmarks · 13
November UK price snapshot
Compare the 16GB RX 9060 XT with the 16GB RTX 5060
Ti if you need a lower price. The same snapshot put them at
roughly £580 and £660, respectively. PC Gamer found
them broadly comparable outside ray tracing, where the GeForce had an
edge. Check Intel’s Arc B580 if you find a substantial
discount and your games run well on it. UK trackers found a roughly £306
listing on 12 November with only one shop showing stock, while PC
Gamer’s later price watch listed one at about £515. That spread is
reason enough to check live stock, compare against Radeon and GeForce
and pay attention to the return policy. If a specific CUDA-dependent
application is part of your work, verify that it runs on the card you
intend to buy before switching. UK
mid-November retail snapshot · PC
Gamer benchmarks and price watch
How to tell them to fuck off
An advert is an invitation to spend money. You can decline. If your
current card still plays the games you bought it for, skip the upgrade.
If you need a replacement, set a price ceiling before you start shopping
and compare real in-stock prices across brands. A discount from last
week’s inflated price is still an inflated price if the performance on
offer falls short of what you can buy elsewhere. Keeping the old card,
buying a sensibly priced used one with buyer protection, or choosing the
rival that offers better value are all ways to refuse the premium.
Who gets to own the means
to use AI?
Are we going back to the Middle Ages? The analogy is about who owns
the resource everyone else needs. If powerful computing is concentrated
in a handful of companies, people with enough money can own it and
decide what gets built. Everyone else increasingly has to rent access on
terms someone else sets. A graphics card getting more expensive is a
small example of what happens when owning the hardware moves further out
of reach.
Follow that logic into education, science or small business. A
university with a large computing budget can afford to run experiments
repeatedly. A small research team on a per-use budget has to decide
which ideas it can afford to test. A wealthy school can buy its pupils
more access. A cash-strapped school may have to make do with whatever
the free tier allows. When the price of participation rises, the richest
can afford to try more things, make more mistakes and learn faster. That
is how a gap in purchasing power can become a gap in who gets to shape
the future.
AI could still deliver enormous public benefits. These deals show how
readily companies will mobilise billions for gains they can capture
themselves. Who is making an equivalent commitment to the independent
researcher, the cash-strapped school or the person who simply wants to
own and run capable hardware at home? Those benefits depend on
deliberate choices about affordability, public access and competition.
Corporate goodwill is a poor substitute for any of them.
The danger goes beyond higher prices. If people are told that an
economic system rewards competition while they watch the most powerful
reserve capacity, finance their own customers and collect public
support, eventually they stop believing the rules are for them. That
discontent can spill into other industries and into politics.
Governments chasing exclusive access to strategic technology can turn
commercial rivalry into geopolitical confrontation. War is a risk when
leaders treat other countries’ access to essential resources as their
own loss. Every step on that path involves human choices.
If AI helps push us towards disaster, the responsibility will lie
with the humans who decide how to use it, who gets access and which
public costs are acceptable in pursuit of private gain. AI gives those
decisions more reach and speed. It gives us more rope. What we do with
it is still up to us.
The public can refuse a bad deal at the checkout. Governments have a
bigger responsibility. If they want people to believe that competition
delivers public benefits, they have to make it possible for people with
less money to participate, build and compete. That is the test.