Margins tell everything

tldr: Margins show whether a company extracts or spreads. Nvidia could cut and still profit, and that cut would flood the market with builders and benefit us all. This is how you judge a company’s altruism.

The choice: extractor or hero

Margins tell everything about the nature of a company, of whether they are purely there for profit extraction or they do carry the mission to make our world better.

How much does a company want to extract? The margins directly show the commitment they have for greater good as opposed to just profits. It is the distance between what we can and what they take.

Do not mistake a company getting 5% margins when market competition allows only that as them being good. Look instead at when a company sells at lower margins than they actually could just so everyone could benefit.

And do not mistake that for purity out of goodwill either. That is also a strategy. This is what I try to point out, this is more of a win-win strategy than anything else – reduce margins to diffuse everything even more.

This came out excessive about one company in one particular market, because it had to be. But the main idea is general.

If a company really cared that technology should penetrate and spread out instead of being concentrated on hook and pay, they would cut margins voluntarily and still get profit.

They did get margins even before the memory price. Call out that they are still using that for profit and profit only, and they behave like now or never. Extract as much now, fuck the world.

Disclaimer: I went too far with actual data when initial point of the article is philosophical one. You can skip to the end or read diagonally.

The cost of manufacture versus the sale

Nvidia never disclosed per-unit cost. No BOM per part, no rack cost – every per-unit number in this piece is a third-party model. Only the margin percentages are audited [1][2]: 72-75% gross, held through 2025 and 2026 straight through the memory spike, easing by guide to no lower than 71-73% next year. Last quarter: $96.2B revenue against $24.1B cost, $89.0B data center; supplier commitments more than doubled in a quarter to $279B [1][2]. They priced the shortage on the way in. The full margin table and revenue-to-cost detail sit in the receipts.

What one unit costs to build – models disagree exactly because nobody publishes [3][4]. A B300 module is around $8,700 of parts at contract, call it $10K with board and packaging slack [3]; no B300 cost on release day was ever published [6][7]. One quoted GB300 superchip figure of $14,104 cannot even hold – two B300 modules alone top that number [12]. The full tree, the HBM contract history and the spot variant sit in the receipts.

What it sells for – no list price exists, and the best models disagree with each other by 1.7x, so any precise rack quote is over-claiming [8][9]. Racks run $3.7M to $6.5M depending on the year and the 15-17% hike [10][11]; per-GPU all-in runs $52-73K street [13][14][15]. The full price table is in the receipts.

Implied component margin starts at 87.7% on H100, walks through 82.1% on B200 and grinds toward ~75% on Rubin Ultra – every generation memory eats a bigger slice of the bill and the curve falls toward the guide [3][9][12]. The component table is in the receipts.

Single B300 – you cannot buy one. Sold on the baseboard, 8 accelerators per board, not sold outside the board. Minimum 2 boards per order [13]. Every margin portrait starts here. Parts at contract: around $8,700, call it $10K with board and packaging slack [3]. Street listing: $50,000-73,000 per GPU through the channel [13][14][15], the same chip sitting at $62,500 late 2025 and $68,750 by April 2026 [15]. Six to eight times the base – and that multiple already held before the memory spike moved the bill.

The number that matters for margins – transfer price $35,000 per GPU. That is the GPU line in the rack ($2.52M) divided by 72 [12]. The $52-69K street number is the same chip marked up by baseboard, switch, CPU, cooling, ODM and channel. The rack is where the real economics live.

And you could not build one yourself – HBM3E is allocation-only, there is no open market [5]. The exact spot-memory arithmetic is in the receipts.

What a B300 actually is – the specs, why the rack is the true product, why the superchip is one coherent pool, and why no card-shaped B300 can exist (watts: 1400W is 2.1x over the highest PCIe power class in existence; design: no NVLink switch domain outside the rack) – all of it lives in the receipts [16][17][18][19][20][21][22][23][24].

How much of the bill is memory: half the build, a tenth of the tag – the share table sits in the receipts [3][5][12].

The cleanest exhibit is the buyable card – RTX PRO 6000 96GB – which walked from $8,565 at launch to $13,250, then $16,000 unannounced, with a card BOM around $3,500-4,500 and every step an implied margin climb from about 50% to about 78% [25][26]. Price history and margin table in the receipts.

Module cost rose worst case about $320 (at $10 per module x32). The card rose $7,435 – so around $7,100 per card is pure margin expansion, sold to the world as a memory story [25][26]. The complaint was never cost-plus: the BOM is around $4K, and $8,565 at launch was already a 2x markup. The complaint is the arithmetic – a $320 memory bill collected as a $7,435 price. Same outrage, no arithmetic hole.

Memory is the broader tax

Memory prices are a separate field which is broader and different – they have product that is used in larger amount of products.

And state the obvious with the cussing it deserves: they are not better. Same extraction, same margins, same playbook. Fuck them too.

But you cannot pin the blame the same way. Memory is a mini-cartel of 3-4 companies, still independent rivals who can undercut each other, versus one player dominating GPUs and holding the keys to everything else – the architecture, the software lock, the switches, the allocation queue. Different animals, same greed.

Memory demand is here forever – every single phone, every single computer, everything now needs more of it every year. Nobody should be under the impression that they are in a market that is going to contract.

That everywhere-point was true a year ago, is true today, will be true in 5 years. Which is exactly why they had more reason than anyone to ramp – and still did not. There is no reason for them not to dump profits into lines right now. And yet everything is booked and nobody can tell you when relief comes.

2026 margins [27][28][29]: SK hynix has beaten Nvidia operating margin for three straight quarters; Micron gross 87% against Nvidia’s 75% (full table in the receipts). The most profitable link in the AI supply chain is no longer the GPU designer.

Contract multiples ran wild within a year – DDR4 up several times over, NAND worse, industry revenue up sharply in a single quarter [28][29][41]. Spot prints, not audited books – the direction is what matters.

The turn: HBM is now less profitable than ordinary DDR5 – HBM around 55-60% margin against around 80% on commodity, and commodity earns multiples of HBM’s gross profit per wafer [30][31][32][33][34]. Why? HBM stacks dies plus packaging, sold on annual contracts that stayed flat while commodity spot ran multiples. So the shortage is not a conspiracy. You make more money stamping ordinary DDR5 than HBM. Supply follows margin, and margin says starve HBM. That is why your B300 costs what it costs, and why your laptop RAM went up too.

Exact law – to hold gross margin flat, price must rise by cost divided by 1 minus margin. At 75% margin, $1 of added cost requires $4 of added price. At 71.5% – 3.51x. At 50% – 2.00x. At 20% – 1.25x. Every dollar the memory maker adds becomes four dollars on a rack – unless Nvidia lets its margin fall, which is exactly what the guide is [2]. They are doing both – hiking 15-17% [10][11] and losing points [2]. It is not a choice. And credit where due – a low-margin Nvidia would be a stabilizer on end prices, the same $1 of cost landing as $1.25 instead of $4. That is true arithmetic.

Four tells for pointing at margins. One – they resell memory at a markup. Rubin bundles modules, markup around 70%. If hyperscalers bought direct, the rack drops $7.8M to $6.7M. The $1.1M difference is the intermediary margin on memory [12]. Note the reversal – on GB300 they shipped only the puck and pushed procurement out to customers, then on Rubin pulled it back in. Margin stacking, on purpose. Two – the 15-17% hike looks bigger than the measured cost delta on GB300 [10][11]. Hike $630-680K per rack, measured HBM content only around $370K, even adding LPDDR you land near $0.5-0.6M total memory content. HBM inflation adds maybe $60K a rack. Pattern of protecting a percentage, not passing a bill. Three – the margin percent is a fraction and its bottom half is exploding. Revenue up 106% with held 75% means absolute cost is soaking the growth [1][2]. Watch the dollars, not the percent. Four – component margin 82% is not company margin 75% because the rack now carries CPUs, switches, cooling, resold memory at thinner economics [12]. That gap is the dilution, and it widens every generation.

But follow the distinction to the end. Nvidia rent is a tax on a narrow capex class – the ten richest buyers on earth, on a product only they buy. Memory rent is a tax on the whole commons, hitting every consumer device on the planet. Samsung memory division made record profit while its own phone and appliance arms posted record losses on its own memory pricing. Apple raised prices, Qualcomm hiked chips, Dell warned. Every HBM stack starves phones, laptops, cars of wafer area. So if the goal is spread across many, Nvidia margin is the less offensive of the two rents. The broad field is the bigger choke. Difference is the population being taxed, not the greed.

And the bust they were reacting to [35][36][37][38][39][40]. 2023: all three of them in record losses at once – a combined $22.6B of damage in twelve months. What they did about it: cut capex hard, wound down wafer starts, cut headcount – and the market applauded, calling the cuts a recovery pulled forward. The exact damage and the exact cuts are in the receipts.

Lag is why this matters. Build plus tool lead plus qualification is 3-5 years, no significant new capacity is likely before 2028, and demand still outruns supply through 2027 [41][42]. The 2022-23 capex cut deleted exactly the capacity that would have landed 2025-27. The shortage is the echo – the numbers in the receipts.

Small correction because it matters – it was not the level of margins. It was the variance. Same firm went plus 15% to minus 24% to plus 72% to plus 76% across four years. A stable 30% would have built far more capacity than a volatile average that swings from minus 24% to plus 76%. Variance is what starves capital, not level. Nobody funds a fab against a distribution with a minus 24% tail when payoff is five years out.

Now my charge, and this part is mine, not the filings. They did not build beforehand when they had the plan and the chance. They knew this was new stuff, memory was going to be demanded in every phone, every laptop, everything – and yet they did not. That was low-risk-mode stupidity, and they own it. They looked at a GPU market dominated by one player extracting massive margins and read it as demand will stay limited to ten buyers – so why build for a thousand? If the market had looked more accessible, if cheaper hardware had signaled a thousand buyers instead of ten, they would have had every reason to fund lines to meet it and extract their own margins. The small visible market gave them the excuse to stay lean.

Wildcard – the cycle may not repeat. Samsung has ~70% of capacity locked through 2031, Micron ~$150B of take-or-pay contracts [5][29]. They are refusing to overbuild right now, which is either discipline or the cartel forming by contract. If contracts replaced the cycle, high margins become permanent rather than spike – and then supply will fix it dies as a hope. The exact contract terms live in the receipts.

What if they cut – 20% and the flood

Price it. On a rack, their slice is around $2.585M against around $568K cost at 78% implied. Cut that to 20% – they charge $0.71M instead of $2.58M. Buyer saves $1,875,320 per rack, $26,046 per GPU. Rack goes $4.00M to $2.12M – 47% cheaper. That is not nothing. That is a genuine enormous act.

One B300 module – at today margin $39,400, at 20% $10,840, at zero $8,670. An 8-GPU node with non-Nvidia content around $157K – today $472K, at 20% $244K down 48%, at zero $226K.

Today they are not even considering because it is too expensive. They are not considering even investing in modernizing existing stuff to just be able to host it. But once you have the signal that prices are low – and that a major player did it – competitors cannot allow themselves to keep high margins anymore. You have a new paradigm where GPU margins are low. That is a signal for everyone to start building more.

Cut margins and invest in lines are not separate. Who says they cannot do it alongside? Come twice as hero. How much they earned – FY23 $27B, FY24 $61B, FY25 $130B, FY26 $216B, this year already $178B in six months guided to $108B next quarter. Insanity either way – but get the right insanity. And FY26 was earned at older pre-hike prices. The 15-17% increases land on systems shipping early next year. So revenue growth is quietly acquiring a we are a memory reseller now component.

What makes us a world where you could buy a B300 with 10% margin? Price it – 78% means rack $4.00M, 20% means $2.12M, 10% means $2.05M, 0% means $1.98M. $1.415M of that rack is not theirs, and it does not care what they decide. So what world is that? Three things have to be true. One – accelerators become interchangeable. Kill the lock and 78% dies within a cycle. In-house silicon is the real undercutter. Two – supply exceeds demand. 10% margins are a bust condition, not a policy. Nobody chooses 10% – markets inflict it. That is what a 10% world looks like on the way in – remember minus 24% in 2023. Three – memory has to break too. Cheaper B300s need the memory bust – forecast for 2028, when the cleanrooms land.

A world where a B300 is sold at 10% margin is a world where capacity exceeds demand. Capacity exceeds demand only because somebody overbuilt. Overbuilding is the diffusion you want. So the cheap B300 is the consequence of diffusion, not the instrument of it.

Once we cut off the prices, nobody can follow with expectancy of high margins anymore. The supply would stay constant. Now you have to consider that one of the competitors might start doing the same. You just need one player, the big one, to set a new precedent, that they voluntarily cut off the margins for greater good.

Jio is the precedent – the same play, documented [69][70]. September 2016: free voice, rock-bottom data tariffs, a market of a billion users [74]. Rivals had little choice but to slash, and demand exploded – 100M subscribers inside 170 days [71]. The full run – year-one market share, subscribers, the first net profit the next fiscal year – sits in the receipts [72][73]. The part nobody predicted: the war ran on for years, and when the survivors finally re-coordinated, tariffs went back up. Price wars destroy capital. Destroyed capital must be recapitalized. Recapitalization requires higher prices.

Samsung proves the first-loss play actually works too. The one firm that refused to cut in 2023 – preemptively invested while bleeding a record loss – and came out holding highest DRAM share since 2016. Your prescription has a precedent, and the precedent won.

Jevons paradox at full power

AI is specifically in very high demand and you could assume that Jevons paradox would work here. So the company can cut the margins and still sell much more, and yet that would drive everywhere to get more products, which allows it to boost everywhere, thus raising demand for the hardware again.

One correction, because the label misleads: a lower price driving more use is just ordinary demand, not the paradox itself. Wider access spreads optimization sideways, and Jevons operates at that scale: efficiency gains increase total demand instead of shrinking it.

One qualification and only one belongs here: under a supply choke the effect is delayed. A price cut today does not fill the queue tomorrow. It is a signal to the market, not an immediate delivery. The full answers to objections live in the cut chapter. Here, just the lag, and it moves on.

Because the rest is not dismissable. Too many forces push the same way at once.

You increasingly run more capability on the same hardware, so hardware does not deprecate as much. Each day brings more optimizations to run better stuff on the same cards, and only the installed memory limits you – and even there the old boxes keep standing. The floor keeps rising under the same metal.

If hardware were accessible then everyone would use it more and think about costs less. It is the same as computers and phones. As they got cheaper, access to the same services equalized, and the internet exploded that way. Nobody asked permission. They just built.

And the current GPUs are already melting trying to serve ever increasing demand. The queue exists before the cut, not after it. So the cut does not invent demand. It uncorks it.

I claim it is a circular cycle that only accelerates in geometric progression. More hands on metal, more optimizations, more capability per unit, more demand for units, more hands on metal. Each turn is faster than the last.

And the important part: it has to be diffusion down to people, not concentration inside big tech. Part of the acceleration is that it is not centralized. Central planning adds raw power every day, more modules, more megawatts – but the center lacks the day-to-day operational knowledge that is only created at local level, in small companies, in labs, in bedrooms. The tricks live where the constraints live.

We already have the evidence. Companies are massively walking away from proprietary software products and moving to what they built themselves. AI penetration happens the same way – from the edges inward, not from the center outward.

Access produces the experiments, and the experiments produce the cheapness. AlexNet 2012 – 2 gaming cards. The field exists because two grad students owned gaming GPUs. FlashAttention 2022 – Stanford lab. vLLM 2023 – Berkeley lab, 2-4x throughput. llama.cpp – individual hobbyists on consumer hardware. Quantization, LoRA – academia plus open source. The two biggest inference-cost wins of the last four years came from university labs. Not Nvidia. Not the hyperscalers. From people who had hardware and time.

I do not prioritize token cost. This thing has different mathematics. Yes hardware grows. My main point is access – as many more people must be able to own as much hardware they want. That is the principle here. Many people operating on hardware start optimizing, start thinking different architecture, lowering the token cost. Only half of the token cost comes from hardware itself and the rest is actually people experimenting on that.

Goal is as most people as possible have more hardware, not concentrated. Then you have natural competition between 1000 groups and 1000 people trying to optimize and lower the cost. And competition between geographical places, where the margin becomes prices, isolation, infrastructure and so on.

And the arithmetic of the goal is pocket-change power, many hands. A 16-GPU box draws about 22 kW. Ten thousand operators with 16 each draw 0.22 GW – a rounding error against one 4.25 GW hyperscaler buildout – but that rounding error is 160,000 GPUs in ten thousand hands.

Fifty thousand operators: 1.1 GW, still pocket change, 800,000 GPUs, fifty thousand owners. The giants’ same silicon, thousands more owners of it. And here is why the point is not tokens. Renting cheap tokens leaves you a customer, and a customer gets repriced whenever the seller feels like it. Owning the hardware makes you a builder: the efficiency of what you own grows because you optimize it, the tools grow because you build them, and the whole industry accelerates the day it stops being ten companies spinning the same ten racks and becomes ten thousand small workshops tuning ten thousand machines. That is what the price cut is actually buying – owners, not bigger bills.

That is the happy cycle. Prices not prohibitive, demand skyrockets again, resource cheaper people want to buy more, people optimize squeeze out more, again more usage.

Measured right now – last gen to this gen price roughly flat to up 10%, throughput up 45%, cost per token down 31%. Next gen price per GPU up 57%, claimed throughput times 3.7, cost per token down 58%. Sticker up 57%. Cost per token down 58%. That is your happy cycle running this quarter. Delivered by more compute per dollar, not the same compute for fewer dollars. To a customer per token those are identical. To the books only one of them pays for the next node.

R&D, monopoly, and who owns the cranes

Ugh but RND and all that. Yes they will claim that they need billions of margins to supply the next area of R&D, but I call this bullshit on a company who were doing this business for more than 20 years. The cost of R&D supposed to get lower and lower and hit some stable threshold. And they already have the very top specialists in the industry already working on them. They have the top expertise. I refuse to believe that they claim they need progressively more and more money for R&D to just ship the same improvements on the constant rate.

The improvements are not even coming that much from architecture – half of them come just from the silicon. You got 1 nm less, you got more density – that is already half of your speedup. I would even claim that very likely if we applied the same architecture we still would be getting an increase without putting a dime into that.

Today they are enjoying the monopoly and having no alternatives and having all the supplies already, the supply and all of market already in their hands beforehand, supported by oppressive and unfair trade regimes and they just continue to extract the margins instead of actually moving humanity forward.

The top AI labs realized it long ago: they cannot rely on a single supplier. Diversifying now, custom silicon as the endpoint. In a couple of years the biggest buyers become the ex-buyers. And what is Nvidia gonna do with that? They already got accustomed to absolutely huge margins – the entire business model is held on that – and their board will not support intentionally risking cutting margins for greater AI diffusion as the top of the market walks. So they sell to the next tier instead – the labs that have not built their own chip architecture yet – at the same decades-old margins. Would the idea of cutting prices to win the rest of the market even arrive? Do you honestly believe they are going to do that, after sitting on the same business model for decades? When faced with market pressure, when losing the advantage, they will try to find some other thing instead of actually increasing the range of customers. And policy hardens in parallel – countries push cloud AI toward national premises, scattering demand instead of piling it on their racks.

CUDA loses position everyday. Everyday software becomes cheaper, so the free world becomes universal. Open weights, portable runtimes, abstracted backends, maturing alternatives. And the volume workload is inference now, far more portable than training.

But point 1 and point 4 cancel each other. CUDA decays daily means the only defense is outspending every rival on software. Sell at cost to save CUDA means defund the only defense. R&D is $28.2B a year. Against gross profit that is 9.8% at 75% margin, 36.7% at 20%, 73.3% at 10%.
Yet again – every month we advance into future when entire software stack will be vibecoded in a week.

What external player is there? This is a complex market. You have many suppliers. TSMC is carrying out the monopoly continued existence, but so it is for everyone else. TSMC is the real chokepoint – and it is neutral. It serves everyone identically, earns from everyone margin, never has to pick a winner. It just rations capacity and re-prices. The shared monopolist.

Memory trio sits at the top of the margin stack with most capacity contracted through 2031. Hyperscalers are buyers, not entrants – they would lose from a price cut. States can fund demand but cannot fabricate the stack. China can – and is fenced off from the frontier by export controls and process limits.

So there is no Jio here. Jio worked because a telecom network is buildable from commodities – towers, spectrum, handsets, all purchasable with money alone. An AI stack needs fab capacity plus HBM plus a software ecosystem. No outsider assembles that with a check. Which means the prescription is not blocked by greed. It just has no one to assign it to. The market has no slot for the actor the argument requires.

The kickstart has never come from the already existing player in this industry. It came from Samsung against Japanese DRAM. It came from Korea against Japan. And it is coming now from Chinese entrants with plus 38% to plus 73% expansions – outsiders with nothing to lose and everything to grab. The entity that would have to eat the first loss is precisely the entity with the most to lose from commoditizing its own product.

The shelf vs the rack

At zero markup the rack is still $1.98M – because 59% of it is memory, board, cooling, power, ODM, CPUs, which they do not set. So margin-shaving buys you a 2.02x price cut. Diffusion needs around 100x. Many does not live behind 47%. It lives behind the BOM.

The saving lands on the least needy link and stops there. Buyers are Microsoft, Google, Meta, Oracle, OpenAI, sovereigns. Do they pass it down? No – cloud rates are set by scarcity, not cost basis. One raised GPU prices 20%, another raised B300 to $9.50 per hour up 21%. So torching $2M per rack of gross profit is hyperscaler margin expansion. You would have made the richest companies on earth 47% richer per rack and charged the rest of us the same.

The lever that actually diffuses is not price. It is rationing – quota racks to universities, national labs, poor countries, outsiders. And that is already happening as a signal. One donated DGX to a naval school with 1,500 students. First tranche in India reserved for national mission applicants, explicitly not walk-in startups. Same compute, different distribution. And it costs zero margin – which is exactly why it is the version that could actually happen. A discount has to be paid for by someone, an allocation just has to be decided.

AWS is the other instrument. Bezos 2005 – when we lower prices we go against the math, volume increase in short term is never enough to pay for price decrease, judgment is that relentlessly returning efficiency improvements and scale economies to customers creates a virtuous cycle. Pricing strategy does not attempt to maximize margin percentages. AWS cut prices 50 times – a 50% reduction every three years. And the part that ends the rack argument: AWS ran 22% and 14% operating margins with analysts estimating 50% plus gross on core products. High volume, high margin business. The most famous price-cutter in corporate history cut 51 times and kept 50% plus gross margins – because the cuts came out of cost declines, not out of margin. Which is what Nvidia is doing. Cost per token down 31% then 58%. Margin held at 75%. AWS playbook, not Jio playbook.

But the door your small bank actually walks through – the products that come in ones. RTX PRO 6000 96GB $8,565 to $16,000 up 87%. DGX Spark 128GB $3,999 to $6,950 up 74%. DGX Station around $93K to $99K to $159K. Mini PC $1,999 to $4,349 up 118%. Every single one is memory-heavy. That tier is where every small bank lives – and memory inflation is what priced them out, not the datacenter margin.

Nvidia’s RTX PRO rise breaks that story in half – the memory excuse covers around $320 a card; the rest, up to $7,100, is their own margin [25][26].

For a company that praises that everyone should get local AI, that the tech should be diffused, that everyone should have access to compute and all that noble crap. And yet look at the practices – by now it is not a rumor anymore, it is a pattern, and I call it a pattern at 95% confidence, because every quarter repeats the same three moves.

They cut the memory. They raise the price. And at the exact moment the market starts signaling that memory prices can hold – they rack the price up again. What a fucking scumbag move.

October 2026 alone, all three at once. DGX Spark – the box promised at $3,000 as Project Digits, and it lands at $6,950. The answer to the outrage: a new affordable $4,999 model with half the memory and half the storage. Half the RAM, and it still costs 25% more than the original 128GB price. You pay more, you get less, and it is sold as accessibility [43]. RTX Spark laptops land the same week – Surface opens at $2,599, and the step from 64GB to 128GB of the same LPDDR5X is priced at $1,600 [47]. The memory tier is the product ladder now.

And the flagship gamer card – killed. October 10: retailers and board partners confirm no new GB202 silicon for GeForce – the chip is reserved for RTX Pro and AI cards [44][45][46]. The 5090 that listed at $2,000 now trades $7,000-9,600 street [49], and the new gaming flagship is a 5080 with 24GB. The everyday people get the leftovers; the pro tier gets the silicon.

The pro tier – the modern Quadro tax. Same silicon, workstation badge, pay 5x to 10x for so called professional cards used for VFX and CAD – mostly you are buying drivers. And that RTX PRO hike landed while spot DRAM was cooling and retail curves were flattening [48][50]. The exact moment the market starts to say memory prices might stay, they rack it up. The scare is the product. The memory is just the excuse.

Spark 128GB of memory is roughly $1,000-1,300 of silicon in a $6,950 box. Cut that margin and a 128GB AI machine lands around $2,800 – and that would genuinely diffuse, tomorrow, to every small bank, lab and shop. No queue. No clean-room lag. It is already on a shelf. A $5,000 96GB card is three times more researchers holding hardware.

And think about what the ladder looks like from below, because for most people cloud is not even the alternative. Many cannot use clouds at all – payment rails they do not have, regions they are not in, or they never find out what is possible because nobody tells them. So they have to run local, or not run anything.

And the API ladder is priced off the top spender: when the top tier costs $50 in API, that is somebody’s weekly salary in much of the world – while that same spend amortizes a GPU at today’s prices in about a year, and at base value in a couple of months.
Owning beats renting far sooner than the gouged tag admits. But the succession path is broken. The ladder should be: cheap first rig, then a bigger one, then one day the big beautiful B300. Instead, gouging starts right at what they call the consumer option – you either show up with ten thousand dollars flat for your first purchase or not using all. The first rig is the expensive one. Cut that barrier – or better subsidize it – and the succession starts working: a player who would never have touched a cloud buys one rig, and the first single GPU unlocks the second. Exactly the Jevons dynamic that runs on tokens – except it catches not just the existing users trying to use more and more, but many normies who never even knew AI could do that.

The rack itself would not diffuse if a bunch of monopolies take everything again (and possibly stockpile).
One might include here the proof of utilization. If someone buys the GPUs, I mean the bigger racks, the customers who buy it en masse, if they cannot prove what they actually utilize it at best capacity, the company should refuse to sell them anymore. But you have the good guy to do that. Nobody but them can enforce that. And if we do, that would be a good choice. That would be another signal that relying on just one monopolist is a bad idea and again helps the entire market.

DeepSeek did the thing

We have a clear example of how DeepSeek are actually following this practice. They release the weights, they discount, and yet they are in profit with already huge enough margins. Consider how such a high percentage margin gets them money and funds them into self-sufficiency to actually provide us with more models in the future. And compare that to how American companies are extracting thousands of percentages from you.

And unlike everyone else in this essay, they proved it in public. February 2025, Open Source Week day six – they published the books of their inference system. Not to analysts, not in filings – a plain post on their own GitHub [51]. One 24-hour window: V3 and R1 served on 278 peak nodes of H800s, about 2,224 GPUs, all-in cost $87,072 per day assuming a $2 per hour lease. Input 608 billion tokens, output 168 billion. If every token had been billed at R1 pricing, daily revenue would be $562,027. Cost profit margin: 545%. In normal accounting that is roughly 84.5% gross – above Nvidia’s 75%.

Now the footnote that should break your brain. Actual revenue, they wrote, is substantially lower. V3 is priced way below R1. Web and app access are free. Off-peak hours get automatic discounts [51]. Nobody made them write that. They published the ceiling of their margin and then confessed they charge below it on purpose.

The margin is real because the inference is real. Cross-node expert parallelism, FP8 from training to serving, computation-communication overlap, load balancers on every phase – and they open-sourced all of it [51]. A single node pushes 73.7k input tokens per second. And this is on H800 – the chip Nvidia sells China castrated by export controls. The same stack runs frontier models on Huawei Ascend silicon – CloudMatrix 384 clusters outperforming Nvidia’s own serving of R1 [53], with R2 rumored trained on Ascend [54]. When your software knows the metal that well, you do not need the monopoly chip, and you do not need the monopoly price.

The what-if did not stay in 2025. The honest part: they did not stay a giveaway forever. April 2026: V4 Preview ships, 1.6T total parameters, 49B active, performance rivaling the world’s top closed-source models, weights open under MIT [57]. Into 2026 they kept running the price war they started – floor pricing, web and app free, automatic off-peak discounts. August 2026: the author of the price war ended it. A significant increase announced [60], then time-based billing restored – peak-hour rates up more than four times from the old flat levels, and on some tiers up to 1,100% [61][62], Reuters confirming both V4-Pro and V4-Flash raised [63]. Bloomberg’s own framing gives the shape away: rates brought closer to rivals’ [61]. Closer, not above.

The receipt for what they could have charged comes from their rivals, not from me. When DeepSeek releases a model, the base API price becomes the standard for the class: Flash landed at $0.60 off-peak, $1.20 peak per million output tokens [58]. Four months later Moonshot ships Kimi K3 – flagship, open weights, same class of release – at $15.00 per million output [64][65][66]. That is roughly four to eight times what DeepSeek charges for the same job, depending on which hour of DeepSeek’s day you catch. Each camp obeys its own founder. OpenRouter resellers sit in a tight band just under Kimi’s launch price, $2.55 in and $12.75 out, DeepInfra at $2.85 and $14.25 [66][67]; nobody races the flagship down to DeepSeek’s floor. Meanwhile DeepInfra turns around and serves DeepSeek’s own Flash at $0.60, the off-peak rate, around the clock [68]. Both directions, same market, the same week. The headroom is on the board for anyone to read: same-class flagship at four to eight times the DeepSeek price. DeepSeek saw the same dashboards. They set their standard on the floor anyway, held it through a year of price war, and when they finally raised, they raised to closer-to-rivals, not into the rival band [61]. Capacity without the grab, printed in public numbers.

Even re-priced, it is efficient as hell. Peak hours are seven of twenty-four, weekdays only – every evening, every weekend, every Chinese holiday runs at half price [58]. Cache-hit input is $0.003 per million tokens. Off-peak Flash output is $0.60. V4-Pro, rivaling the top closed-source models, tops out at $3.96 in peak hours [58]. Today’s markup sits around 60% over serving cost against the 545% they once published [my claim]. They never hid the first number, and they are not hiding the second – meanwhile the frontier collapsed onto the floor they set. o1 charged $60 per million output when R1 landed [56]. GPT-6 starts at $0.50 today [59]. A 120-fold fall in under two years, and V4.1-Flash scores a point behind Luna overall while beating it on coding and agentic lanes at $1.20 peak [59]. That is diffusion at model level – a price war, and the same lab ending it. Still in profit on every token it sells. That is why they are the heroes.

Just cutting the margin does not solve everything

The cut is the kick that sets the market in motion even if nothing happens for a while. A couple of problems first.

The same 10 buyers take it all

Nvidia would just more likely sell to the same 10 data centers they were selling already, because the order is much bulkier than newly formed customers.
Fix is price cut plus allocation rules. A price cut alone does not fix allocation. Pair it with capped orders per buyer or reserved batches for new customers. Say that out loud, and you look like you thought it through.
But nobody would check up on it. Who can enforce it? Nobody. The safeguard is practical, not contractual – buyers cannot finance much more than they can host within a year anyway.

Once more the objection: those companies could just keep buying anyway. If they were ready to pay 3x, they can pay 1x to hold for later. Warehouses full, utilization someday.

Hoarding hardware is expensive and risky. Chips depreciate fast. Next-gen arrives and the warehouse stock is last-year model. Stockpiling with no power or rack space to run it is dead capital, and a data center that cannot host it has no reason to buy it. And you cannot say they are both swimming in spare hosting room and hoarding because they cannot use it all. Pick one.

Power and premises cap the giants

Yet it is also the objection that the customers in question are limited by power – not enough electricity – and by the actual capacity to host the hardware in question. So you have the opposite force: they cannot actually buy it all even if they wanted to. And it might actually diffuse into data centers in different countries that already have the capacity to host but just not able to because of prices.

And if the prices were actually affordable, many countries – many data centers – with favorable conditions but without capital would start improving their infrastructure right away.
This one actually helps. If the giants physically cannot absorb unlimited chips, then supply overflows to the next tier – smaller data centers, other countries, labs with spare power but no budget. Price was the only thing keeping them out.

Demand is a quantity ceiling, not a price ceiling. Around 1.5M GPUs per 4.25 GW – 353 GPUs per MW. What they can swallow is measured in gigawatts, not dollars. They would take more at any price and cannot, because the shell and the substation are not there. Guaranteeing land, power and shell for $105B cap is precisely because GPUs have nowhere to sit.

Per gigawatt at 140 kW per rack – 7,143 racks, $28.6B of gear. At today margin content per GW $28.5B, at 20% margin $15.2B. Down 47% on the single biggest line item of a gigawatt. Total build per GW maybe down 45%. In a country where capital is the binding constraint and power is cheap, that is not a discount – it is the difference between hosting and not hosting.

But cheap-power places have a different gate. Cheap electricity is not a live grid connection. Substations, water, fibre, reliability. Capital and power bind at once – which is exactly why the buyer of record there is the state. States have capital and no margin to cut. The places with cheap power mostly lack grid, capital and stability. The places with capital have constrained, expensive power. The world does not have a cheap-power-plus-capital country at scale. That is a grid problem before it is a GPU problem. Cheap silicon does not build a substation – the substation is what the small bank is actually waiting for.

Neutral now, pays when scarcity breaks

Would I be able to say that currently it cancels out, so it is neutral? You get lower prices, yet existing customers will buy it all anyway, just stockpile until they have capacity. But that is until you have the severe choke on actual supply. Once scarcity becomes less, the companies hold hard, immediately weighting the risk, thus getting out of books and allowing tier 2 to spread.

Right now with supply severely choked, the price cut is roughly neutral. Existing buyers absorb everything and stockpile. Nobody new gets in. Fine, concede that. Neutral is a snapshot. Scarcity is temporary, and the moment supply loosens the logic flips. Hoarding stops making sense when you can just buy on demand. Holding depreciating chips becomes pure risk. The big buyers dump or stop stockpiling, and that inventory and allocation spills down to the next tier. At that point the lower price is what lets the new customers actually afford to step in.

So the price cut is not a fix for today. It is a bet on the moment scarcity breaks. Under high margins, even when supply loosens, newcomers are still priced out and the market stays small. Under lower margins, they are ready to flood in the second the door opens.

The answer: neutral in the short run, positive once supply loosens, not nothing. The timing objection is obvious – if it is neutral now, why cut now? Wait until supply loosens. Because the cut is what makes supply loosen: it pushes the memory makers to build. The cut is not a reaction to the thaw. It is the thing that melts the ice.

And the hard math underneath: new supply minus what the giants can swallow is still below zero. A discount against a hole like that just reshuffles the queue – it does not diffuse anything. The rule is correct, it just fires at the same 2028 door.

Be the hero

My whole idea was that you need one to start to kick the bucket.

But I argue they could come out as heroes if they start cutting the margins and getting more players able to actually participate. That shakes the entire industry also. When prices are not prohibitive, demand skyrockets again and that also partially helps the driven products.

They have this choice, but they are not making it. And they go down in history as a company that choked the AI developers. They choked on the demand that everyone desperately wanted. And instead just sucked money dry of everyone. Instead of actually being the good company that ushers us into the new era.

The diffusion you want arrives anyway – by accident, not by decision. The 2028 wave. New entrants coming in from outside. Domestic silicon serving a billion people because it has to. Cheaper memory normalizing and dragging the card tier down with it. Nobody decides cheapness. It shows up when the fabs land.

Rule stands same everywhere

I wanted it to be a more philosophical article of how the company can be judged by the difference of what market allows to extract, whether they voluntarily limit themselves for the collective good.
I carried away a little with facts and calculations, which many of them would be obsolete in the very next week, possibly. And yet the core point still stands.

When the market is tight, competitive or rather healthy, when you have competition and everyone unable to extract more than 5-10% of margins, you could not really blame them for being good and bad. They have to operate in such circumstances. But when you have a monopoly, when you have little players, by amount, when you have a situation where somebody can get huge amounts of money by extorting prizes, by raising them, that just shows what they truly care about.
And yet I showed an explanation how sometimes when demand is high they can reduce margins and still let everyone win, but the diffusion is more win-win scenario even for existing monopoly.

Just notice the next time they claim what they are doing good, ask them – why dont you try drop prices then?

The receipts

For more detail and actual data – move it here as reading the whole stack in the main body became painful.

Nvidia margin detail [1][2]

Period	Gross margin
Q2 FY26 (Jul 2025, GB300 shipping)	72.4%
Q3 FY26 (Oct 2025)	73.4%
Q4 FY26 (Jan 2026)	75.0%
Q1 FY27 (Apr 2026)	74.9%
Q2 FY27 (Jul 2026)	75.0%
Q3 FY27 guide	74.0%
Q4 FY27 guide	71-72%
FY28 guide	72-73%
Full FY25	75.0%
Full FY26	71.1%

Cost of revenue FY26 was $62.475B, up 91.4% while revenue went up 65.5%. Last quarter revenue $96.221B against $24.079B cost. Revenue up 105.9%, cost up 86.8% – cost still growing slower than revenue. Data Center alone $89.0B [1][2]. CFO Kress on the call – extreme pricing conditions in memory, magnitude exceeded prior expectations and headed even higher into next year [2]. Supplier commitments more than doubled in a quarter to $279B, inventory up to $32B [1][2].

One unit: the models that disagree [3][4][5][6][7]

Part	B200 reference (Epoch, Dec 2025)	B300 now (derived, Aug 2026 contract)	B300 at spot
HBM	$2,900 (192 GB)	$5,040 (288 GB at $17.5/GB)	$16,799 (at $58.33/GB) [5]
Logic dies, 2x ~800 sqmm on 4NP	$900	$900	$900
CoWoS-L packaging	$1,100	$1,250	$1,250
Packaging yield loss	$1,000	$1,000	$1,000
Board, assembly, test	$480	$480	$480
Total module	~$6,250, range $5,700-7,300	~$8,670	~$20,429

HBM was $4,464 at old contract ($15.5/GB, Dec 2025) and $5,040 at new contract ($17.5/GB, Aug 2026) [3][5]. That is the only rigorous public model, and it is for the previous chip. B300 with 96 GB more memory derives to roughly $7,700-8,670 depending on packaging assumptions – use the range, not the point.

One industry model says GB300 superchip (2 GPUs plus 1 Grace) is $14,104 [12]. That cannot hold if Epoch is right, because two B300 modules alone are $15.2-15.4K plus Grace die plus 512 GB LPDDR5X. True number is probably $17-19K.

What it sells for [8][9][10][11][12][13][14][15]

Unit	Then	Now
GB200 NVL72 rack	$3.1-3.9M (SemiAnalysis via [9])	–
GB300 NVL72 rack	$3.7-4.0M, Mar 2025 Loop Capital [8][9]	$4.3-4.7M after 15-17% hike on systems shipping early 2027 [10][11]
Inference-tuned GB300 rack	–	$6.0-6.5M, Mar 2026 supply-chain [9]
Per GPU all-in	$51.4-55.6K implied [9]	$62.5K late 2025 to $68.75K Apr 2026 street [15], $52-73K range in listings [13][14]
Rubin VR200 rack	$7.8M	$8.1-9.1M after hike [10]
Rubin GPU / Vera CPU in volume	–	$55,000 / $5,000 [12]

Loop Capital versus Tom’s supply-chain are 1.7x apart [9]. The 15-17% hike hits Grace Blackwell and Vera Rubin [10][11]. Applied to a gigawatt of build it is plus $5B minimum.

Component-margin curve [3][9][12]

Generation	BOM versus sale	Implied margin
H100	–	87.7%
H200	–	86.4%
B200	–	82.1%
GB200 superchip	$14.2K / $65K	78.2%
GB300 superchip	$14.1K / $80K	82.4%, accretive, Kress confirmed improved mix from Blackwell Ultra [2]
Rubin superchip	$27.5K / $130K	78.8%
Rubin Ultra	$42.2K / $168.6K	~75.0%

Rack BOM goes $4.0M to $7.8M next gen, memory share under 10% to 26% [12]. What the guide costs them – on around $108B a quarter, 75% to 71.5% is minus $3.8B gross profit per quarter, around $15B a year. Calculation, not disclosure.

Spot HBM [3][5]

Pure components for one B300 at contract are around $8,670. At spot the memory alone ($16,799) costs 1.94x more than the entire module BOM at contract [3][5]. The $2,100 stack price is a scarce-lot report, not something you can order [5].

What a B300 is [16][17][18][6][7]

	B200	B300 Blackwell Ultra
HBM3E per GPU	192 GB	288 GB, 12-hi stacks, the real flex
Dense NVFP4	–	15 PFLOPS, 1.5x B200, ~2x attention
Power per GPU	–	1200W to 1400W configurable
Networking	–	ConnectX-8, 800 Gb/s per GPU, 2x B200
Silicon	–	208B transistors, TSMC 4NP

The rack is the actual product [9][17] – 72 GPUs plus 36 Grace CPUs, 20 TB HBM, 132-142 kW nominal around 155 kW peak, liquid-cooled, $3.7-4.0M a rack, ~1,580 kg. That is 15x a normal enterprise rack – power and cooling are the bottleneck now, not chips.

The superchip is not two B300s. It is one Grace CPU and 2 B300 GPUs behind a single coherent address space, about 1 TB of memory shared over NVLink-C2C at 900 GB/s [16][17] – a GPU kernel dereferences host pointers directly, so LPDDR works as a slower tier of GPU memory instead of a wall.

Not new. Superchip pattern is 2.5 years old. Grace Blackwell Superchip is GTC Mar 18, 2024. Ultra version announced GTC Mar 18, 2025 [18]. First rack Jul 3, 2025, Dell to CoreWeave, literal first one on earth [6][7]. Volume fall 2025, mass production for about a year. Vera Rubin went into production Sep 30, 2026. GB300 is now the value tier, not the frontier.

Why no card-shaped B300 [19][20][21][22][23][24]

No PCIe B300, and it is not the package. H100 PCIe (350W, 80 GB HBM on the same interposer, socketed card) [19] and H200 NVL (141 GB HBM3e, PCIe 5.0 x16, 600W, identical capacity to H200 SXM) [20] prove HBM in a slot was always possible. The blocker is watts. PCIe CEM tops at 300W standard levels with 600W max on 12VHPWR and 675W highest supported in CEM 5.1 [21]. B300 is 1400W, 2.1x over the highest PCIe power class in existence. Second blocker – PCIe guts the product. NVLink is 7x PCIe Gen6 x16, no switch domain means no 72-GPU shared memory and no coherent pool to Grace [23]. Dell measured 128 tok/s SXM versus 98.8 tok/s PCIe, 30% just from form factor [24]. Nvidia said it out loud – B200 in PCIe form are not yet available, B100 not in production, they never became available, buy RTX Pro PCIe cards with the same die but without HBM instead [22]. Scarce HBM goes into $4M racks, not into one-per-customer cards. The card stopped being the profitable shape.

How much of it is memory [3][5][12]

Where	Memory inside	Box cost / price	Memory share
B200 module BOM (Epoch)	$2,900 HBM of ~$6,250	build cost	~46%
B300 module BOM at contract	$5,040 HBM of ~$8,670	build cost	~58%
B300 module BOM at spot	$16,799 HBM of ~$20,429	build cost	~82%
GB300 rack	~$370K HBM, ~$0.5-0.6M with LPDDR, of ~$4.0M rack	sale price	under 10% then, 26% next gen [12]
RTX PRO 6000 at $16,000 tag	$1,920-2,240 GDDR7 (32x 3GB at $60-70)	sale price	12-14%
RTX PRO 6000 build	same $1,920-2,240 of ~$3,500-4,500 BOM	build cost	roughly half

Read it as: memory is half the build and a tenth of the tag. The rest is margin stack.

RTX PRO 6000 exhibit [25][26]

When	Price	Rise	Memory inside
Apr 2025 launch	$8,565	–	$1,920-2,240
Jun 2026	$13,250	+$4,685, +55%	same 32 modules
Aug 12 2026, unannounced	$16,000	+$7,435 total, +87%	same 32 modules
Street now	$15,499-17,999	list $17,999 at top	same 32 modules

Card BOM is roughly $3,500-4,500 (modules plus ~750 sqmm die, board, 600W cooler – estimate).

Price	Implied gross margin on $3.5-4.5K BOM
$8,565	~50-55% – already a 2x markup at launch
$13,250	66-74%
$16,000	72-78%
$17,999 list	75-81%

Memory makers 2026 [27][28][29]

Maker	Operating margin	Gross margin
Micron	82.3% (FY26 Q4, full year 81.1%) [29]	87.0% [29]
Samsung DRAM	~78.5-80% (division carried whole company) [28]	–
SK hynix	76.3% (Q2 2026) [27]	–
TSMC	60.3% (estimate)	–
Nvidia	66.2% (derived from [1][2])	75.0% [1][2]

HBM versus commodity, source by source [30][31][32][33][34]

SK chairman: HBM ~60%, standard ~80%. Micron CEO: non-HBM margins currently higher than HBM. JP Morgan: HBM ~55% vs commodity ~81%. Bernstein: per wafer commodity earns ~2x revenue and ~3x gross margin of HBM. HBM4 eats roughly a quarter of all DRAM wafer capacity.

The 2023 bust, exact damage [35][36][37][38][39][40]

SK hynix 2023 revenue down 27%, operating loss with margin minus 24% against profit the year before. 39-point swing in twelve months. Record quarterly losses. Samsung chips record loss from profit the year before. Whole-company worst since 2008. Micron revenue halved, net loss $5.83B, free cash flow minus $5.45B. Combined damage around $22.6B in one year.

What they did – SK hynix halved capex, Micron cut capex ~40%, wafer spend down half, wafer starts down 20% then approaching 30%, headcount down 15%, bonuses suspended [36][38][39]. And the market rewarded not building – analysts called the capex cut positive, it would pull the recovery forward.

The lag numbers [41][42]

New capacity delayed by equipment, construction, long cycles – phased starting 2027. Clean-room constraint – no physical floor space. Server DRAM plus HBM bit supply up ~27% in H2 2027 [42], memory share of cloud capex 47% in 2026 toward 68% in 2027 [41].

Wildcard contract terms [5][27][29]

Micron 26 take-or-pay contracts ~$150B, even at floor prices margins above any prior cycle peak [29]. SK hynix long contracts with ~10 customers, CapEx Discipline at 76% operating margin [27]. Bookings run through late 2026; the earliest surplus forecasts point at 2028.

Jio receipts [71][72][73]

6.4% market share in year one, 72.16M subscribers [73]. 186.6M subscribers with first commercial-year net profit by March 2018 [72].

Sources

[1] Nvidia Q2 FY27 results – https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
[2] Q2 FY27 CFO commentary (PDF) – https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/Q2FY27-CFO-Commentary.pdf
[3] Epoch AI B200 cost breakdown – https://epoch.ai/data-insights/b200-cost-breakdown
[4] Epoch chip sales methodology – https://epoch.ai/data/ai-chip-sales-documentation/methodology
[5] TrendForce HBM3E spot 4-5x LTA – https://www.trendforce.com/news/2026/09/01/news-hbm3e-spot-prices-said-to-be-4-5x-lta-levels-with-samsung-reportedly-locking-70-of-memory-capacity-through-2031/
[6] Dell first GB300 NVL72 to CoreWeave, Jul 3 2025 – https://www.dell.com/en-us/blog/dell-delivers-market-s-first-nvidia-gb300-nvl72-to-coreweave/
[7] CoreWeave first GB300 NVL72 deployment, Jul 3 2025 – https://www.coreweave.com/news/coreweave-becomes-first-cloud-provider-to-deploy-nvidia-gb300-nvl72-platform
[8] DCD on Loop Capital $3.7-4.0M per rack – https://www.datacenterdynamics.com/en/news/apple-to-spend-1bn-on-nvidia-gb300-nvl72-systems-report/
[9] GPUAdvisor GB300 pricing reality check – https://gpuadvisor.com/gb300-nvl72-pricing
[10] Bloomberg server price hikes above 15 percent, Aug 22 2026 – https://www.bloomberg.com/news/articles/2026-08-22/nvidia-customers-notified-about-ai-related-price-hikes-above-15
[11] Seoul Economic Daily 17 percent – https://en.sedaily.com/international/2026/08/23/nvidia-to-raise-ai-server-prices-by-up-to-17-percent
[12] Morgan Stanley Rubin BOM teardown via locsic – https://locsic.com/thinking/morgan-stanley-rubin-bom-teardown/
[13] Gatewell B300 baseboard pricing – https://gatewellusa.com/ai-compute/p/nvidia-b300-sxm
[14] Sonic Components B300 list – https://www.soniccomponents.com/product/935-26287-0070-000-nvidia/
[15] Reuters China-US B300 server price, Apr 30 2026 – https://www.reuters.com/world/china/prices-nvidias-b300-server-1-million-china-us-curbs-sources-say-2026-04-30/
[16] Nvidia Inside Blackwell Ultra dev blog – https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/
[17] nvl72.dev compute tray breakdown – https://nvl72.dev/compute-tray
[18] Tom’s Hardware B300 announcement, Mar 18 2025 – https://www.tomshardware.com/pc-components/gpus/nvidia-announces-blackwell-ultra-b300-1-5x-faster-than-b200-with-288gb-hbm3e-and-15-pflops-dense-fp4
[19] H100 PCIe product brief – https://www.nvidia.com/content/dam/en-zz/Solutions/gtcs22/data-center/h100/PB-11133-001_v01.pdf
[20] Heise H200 NVL PCIe 600W 141 GB – https://www.heise.de/en/news/Nvidia-H200-Hopper-also-available-as-a-PCIe-card-10081290.html
[21] PCI-SIG CEM 5.1, 675W max class – https://pcisig.com/PCIExpress/Specs/CEM/CardElectromechanical_5.1_1.0
[22] LeMagIT Salvator interview on PCIe tier – https://www.lemagit.fr/actualites/366621277/GTC2025-Nvidia-recentre-sa-production-sur-les-GPU-haut-de-gamme
[23] packet.ai SXM versus PCIe – https://packet.ai/blog/sxm-vs-pcie-gpus
[24] Dell SXM versus PCIe benchmark – https://infohub.delltechnologies.com/en-us/p/comparing-nvidia-h100-pcie-on-r760xa-vs-nvidia-h100-sxm-on-xe9680/
[25] TechPowerUp RTX PRO 6000 memory math – https://www.techpowerup.com/351549/nvidia-rtx-pro-6000-blackwell-96-gb-gpu-now-costs-usd-16-000
[26] Tom’s Hardware RTX PRO 6000 $16K MSRP – https://www.tomshardware.com/pc-components/gpus/nvidia-doubles-rtx-pro-6000-blackwells-msrp-to-a-staggering-usd16-000-96gb-card-started-pre-orders-below-usd8-000-last-year
[27] SK hynix Q2 2026 results (76% op margin, CapEx Discipline, 10 LTAs) – https://news.skhynix.com/en/q2-2026-business-results/
[28] Seoul Economic Daily memory margins top Nvidia – https://en.sedaily.com/finance/2026/10/09/samsung-memory-margins-top-nvidia-as-profit-outlook-soars
[29] Micron FY26 Q4 (87% gross, $150B RPO) – https://investors.micron.com/news/press-release/2026/Micron-Technology-Inc–Reports-Record-Fiscal-Fourth-Quarter-and-Full-Year-2026-Results/
[30] TrendForce DDR5 margins surpass HBM – https://www.trendforce.com/news/2026/03/23/news-while-ddr5-margins-reportedly-surpass-hbm-whats-next-for-memory-giants/
[31] Bernstein 2x revenue 3x margin per wafer via techflowpost – https://www.techflowpost.com/en-US/article/32157
[32] UBS Micron DRAM vs HBM via vgtimes – https://vgtimes.com/tech-and-hardware/164184-microns-ordinary-dram-is-more-profitable-than-hbm-analysts-say.html
[33] Noah News Micron profits driven by conventional DRAM – https://noah-news.com/microns-profits-increasingly-driven-by-conventional-dram-over-high-bandwidth-mem/
[34] JP Morgan HBM vs DRAM margins via twiscan – https://twiscan.com/en/x/MojoTricks/2106252994102898872
[35] SK hynix record loss Jan 2023 Bloomberg – https://www.bloomberg.com/news/articles/2023-01-31/hynix-reports-biggest-loss-on-record-on-memory-slump
[36] SK hynix Q4 loss Reuters Jan 2023 – https://www.reuters.com/technology/sk-hynix-reports-q4-loss-demand-weakens-2023-01-31/
[37] Samsung Q4 and FY2023 results – https://news.samsung.com/global/samsung-electronics-announces-fourth-quarter-and-fy-2023-results
[38] Micron FY2023 results – https://investors.micron.com/news/press-release/2023/Micron-Technology-Inc–Reports-Results-for-the-Fourth-Quarter-and-Full-Year-of-Fiscal-2023-09-27-2023/default.aspx
[39] Reuters Micron to supply fewer chips Nov 2022 – https://www.reuters.com/technology/micron-expects-supply-fewer-memory-chips-2023-2022-11-16/
[40] Reuters Samsung Q4 profit falls Jan 2024 – https://www.reuters.com/technology/samsung-electronics-q4-profit-falls-34-amid-weak-consumer-demand-2024-01-30/
[41] TrendForce memory share of CSP capex to 68 percent – https://www.trendforce.com/presscenter/news/20260825-13198.html
[42] TrendForce 2027 DRAM structural tightness – https://www.trendforce.com/research/download/RP260727PO
[43] The Register – DGX Spark $4,999 half-RAM SKU, 128GB jacked to $6,950 – https://www.theregister.com/systems/2026/10/02/nvidia-debuts-4999-dgx-spark-with-half-the-ram-and-storage-amid-memory-crunch/5300622
[44] VideoCardz – GB202 reserved for RTX Pro, 5090 discontinued – https://videocardz.com/newz/nvidia-reportedly-discontinuing-rtx-5090-gb202-gpus-to-be-reserved-for-rtx-pro-series
[45] PC Games Hardware – 5090 EOL confirmed by retailers and partners – https://www.pcgameshardware.de/Geforce-RTX-5090-Grafikkarte-281029/Specials/Price-End-of-Life-RTX-5080-24GB-Release-1556889/
[46] Tom’s Hardware – RTX 5090 production halt, 5080 24GB as new flagship – https://www.tomshardware.com/pc-components/gpus/nvidia-reportedly-halts-geforce-rtx-5090-production-in-favor-of-ai-data-center-and-professional-gpus-impending-supply-drought-expected-to-drive-up-prices-rtx-5080-24gb-rumored-as-new-gaming-flagship
[47] TweakTown – Surface Laptop Ultra $2,599, 64GB to 128GB step costs $1,600 – https://www.tweaktown.com/news/113971/microsoft-opens-surface-laptop-ultra-preorders-at-dollars2599-with-rtx-spark-n1x-inside/index.html
[48] WindowsForum – RTX PRO 6000 $13,250 to $16,000, +20.8% in two months – https://windowsforum.com/news/rtx-pro-6000-blackwell-hits-16-000-why-nvidias-96gb-gpu-costs-far-more-than-rtx-5090.446146/
[49] Gate News – 5090 street price $7,000-9,600 – https://www.gate.com/news/detail/NVDA/a-leaker-claims-the-rtx-5090-may-have-been-discontinued-with-market-prices-24939763
[50] RAM Prices USA – DDR5 tracker, 30-day +5.9% vs 90-day +19.9% – https://rampricesusa.com/ddr5-ram-prices
[51] DeepSeek open-infra-index, Day 6 inference system overview (545% margin disclosure) – https://github.com/deepseek-ai/open-infra-index/blob/main/202502OpenSourceWeek/day_6_one_more_thing_deepseekV3R1_inference_system_overview.md
[52] Reuters – DeepSeek claims theoretical cost-profit ratio of 545% per day – https://www.reuters.com/technology/chinas-deepseek-claims-theoretical-cost-profit-ratio-545-per-day-2025-03-01
[53] SCMP – Huawei Ascend CloudMatrix 384 outperforming Nvidia serving R1 – https://www.scmp.com/tech/tech-trends/article/3315068/how-huaweis-ascend-ai-chips-outperform-nvidia-processors-running-deepseeks-r1-model
[54] Huawei Central – R2 rumored built on Ascend – https://www.huaweicentral.com/deepseek-rumored-to-build-r2-ai-model-using-huawei-ai-chips
[55] DeepSeek API docs – V3.2-Exp with 50%+ price cuts – https://api-docs.deepseek.com/news/news250929/
[56] OpenAI API pricing – o1 $60/M output tokens – https://openai.com/api/pricing/
[57] DeepSeek API docs – V4 Preview launch, 1.6T/49B open-sourced – https://api-docs.deepseek.com/news/news260424/
[58] DeepSeek official pricing – V4-Pro and V4.1-Flash, peak/off-peak tiers – https://api-docs.deepseek.com/quick_start/pricing/
[59] BenchLM – DeepSeek V4.1 Flash vs GPT-6 Luna, benchmarks and pricing – https://benchlm.ai/compare/deepseek-v4-1-flash-vs-gpt-6-luna
[60] Bloomberg Aug 6 2026 – DeepSeek plans significant price increase for its AI services – https://www.bloomberg.com/news/articles/2026-08-06/deepseek-plans-significant-price-increase-for-its-ai-services
[61] Bloomberg Aug 13 2026 – DeepSeek increases prices by multiple times, peak-hour rates 4x+ – https://www.bloomberg.com/news/articles/2026-08-13/deepseek-increases-prices-for-ai-services-by-multiple-times
[62] Quartz via Yahoo Finance – API prices up to 1,100% starting Aug 16 – https://finance.yahoo.com/technology/ai/articles/deepseek-raising-api-prices-1-174027670.html
[63] Reuters Aug 13 2026 – V4-Pro and V4-Flash API pricing raised – https://www.reuters.com/world/china/deepseek-raises-api-pricing-its-v4-models-2026-08-13
[64] BenchLM – Kimi API pricing, K3 at $3 in / $15 out – https://benchlm.ai/moonshot/api-pricing
[65] Puter developer guide – Kimi K3 released July 16 2026 at $3/$15 – https://developer.puter.com/tutorials/kimi-api-pricing
[66] OpenRouter – Kimi K3 model page, $2.55 in / $12.75 out, 2.8T open-weight – https://openrouter.ai/moonshotai/kimi-k3
[67] DeepInfra – Kimi-K3 API, $2.85 in / $14.25 out – https://deepinfra.com/moonshotai/Kimi-K3/api
[68] DeepInfra – DeepSeek-V4.1-Flash served at $0.60/M output around the clock – https://deepinfra.com/deepseek-ai/DeepSeek-V4.1-Flash
[69] RIL media release, Jio launch Sep 1 2016 – free voice Welcome Offer – https://www.ril.com/sites/default/files/2023-01/Media-Release-Jio-01092016-1.pdf
[70] Business Standard – Ambani announces free voice, cheapest data tariffs – https://www.business-standard.com/article/companies/reliance-jio-to-offer-free-voice-calls-cheaper-data-tariffs-mukesh-ambani-116090100489_1.html
[71] RIL media release Feb 21 2017 – 100M subscribers in 170 days – https://www.ril.com/sites/default/files/2023-01/Media-Release-JIO-21022017.pdf
[72] Jio media release Apr 27 2018 – FY18: 186.6M subscribers, first commercial-year net profit – https://myjiostatic.cdn.jio.com/cms/assets/pressrelease/Jio-Media-Release-4Q+FY-17-18.pdf
[73] TRAI year-end 2016 share data via Economic Times – Jio at 72.16M subscribers, 6.4% share – https://economictimes.indiatimes.com/news/company/corporate-trends/reliance-jio-eats-into-market-share-of-bharti-airtel-vodafone-other-large-telcos-in-2016-trai/articleshow/59428819.cms
[74] The Guardian – Jio launches cheap 4G, price war in a billion-user market – https://www.theguardian.com/world/2016/sep/05/reliance-jio-launches-cheap-4g-services-in-india