[HN Gopher] Nvidia is gearing up to sell servers instead of just...
       ___________________________________________________________________
        
       Nvidia is gearing up to sell servers instead of just GPUs and
       components
        
       Author : giuliomagnifico
       Score  : 159 points
       Date   : 2025-11-14 13:18 UTC (9 hours ago)
        
 (HTM) web link (www.tomshardware.com)
 (TXT) w3m dump (www.tomshardware.com)
        
       | re-thc wrote:
       | Soon Nvidia will sell AI itself instead of servers.
        
         | michaelbuckbee wrote:
         | Considering they have a pure service in selling Geforce Now
         | (game streaming), that doesn't seem in any way far fetched.
        
         | Cthulhu_ wrote:
         | To a point / by some definitions of the phrase AI they already
         | do: https://en.wikipedia.org/wiki/Deep_Learning_Super_Sampling
         | 
         | I wouldn't be surprised if we see some major acquisitions or
         | mergers happening in the next few years by one of the
         | independent AI vendors like OpenAI and Nvidia.
        
         | Palomides wrote:
         | why? selling GPUs is way more profitable
        
           | reactordev wrote:
           | Why sell when you can rent?
        
             | Palomides wrote:
             | so you don't get stuck with many billions of dollars in
             | useless GPUs and data centers when the bubble pops
        
             | re-thc wrote:
             | That's what Coreweave etc does and Nvidia already invests
             | in them (i.e. has a stake).
        
           | giuliomagnifico wrote:
           | Sell a whole infrastructure is more profitable than sell
           | components, and also put the customers in "sandboxes" with
           | you.
        
       | hvenev wrote:
       | Don't they already sell servers? https://www.nvidia.com/en-
       | us/data-center/dgx-platform/
        
         | p4ul wrote:
         | I had the same reaction. Haven't they been selling DGX boxes
         | for almost 10 years now? And they've been selling the rack-
         | scale NVL72 beast for probably a few years.[1]
         | 
         | What is changing?
         | 
         | [1] https://www.nvidia.com/en-us/data-center/gb200-nvl72/
        
           | AlanYx wrote:
           | When nVIDIA sells DGX directly they usually still partner
           | with SuperMicro, etc. for deployment and support. It sounds
           | like they're going to be offering those services in-house
           | now, competing with their resellers on that front.
        
           | reactordev wrote:
           | Cutting out the Vendor like SuperMicro or HPE, they are going
           | straight to consumer now.
        
         | rlupi wrote:
         | Hyperscalers and similar clients don't use DGX, but their own
         | designs that integrate better with their custom designed
         | datacenters
         | 
         | https://www.nvidia.com/en-us/data-center/products/mgx/
        
       | lvl155 wrote:
       | We're not far from Nvidia exclusively bundling ChatGPT. It's a
       | classic playbook from Microsoft.
        
         | MangoToupe wrote:
         | Why would Nvidia ever agree to that?
        
         | ptero wrote:
         | Chatgpt is not the only game in town. Any exclusivity deal will
         | likely backfire against chatgpt.
        
         | mcintyre1994 wrote:
         | I'm pretty sure they'd like to keep selling chips to all of
         | OpenAI's competitors too.
        
         | gruturo wrote:
         | ChatGPT doesn't really have much of a moat. If it becomes
         | Microsoft or Nvidia exclusive, it just opens an opportunity for
         | its competitors. I barely notice which LLM I'm using unless
         | it's something super specific where one is known to be
         | stronger.
        
       | dmboyd wrote:
       | Aren't they already supply constrained? Seems like this would be
       | counterproductive in further limiting supply vs a strategy of
       | commoditizing your complements. This seems closer to PR designed
       | to boost share price rather than a cogent strategy.
        
         | MattRix wrote:
         | They're only supply constrained on the chips themselves.
         | Selling fully integrated racks allows them to get even more
         | money per chip.
        
         | dboreham wrote:
         | In MBA-speak this is "capturing more of the value chain".
        
         | mikeryan wrote:
         | Huh. I view it the other way. If you're supply constrained go
         | straight to the consumer and capture the value that the
         | middlemen building on top of your tech are currently profiting
         | from.
        
         | energy123 wrote:
         | > Further limiting supply
         | 
         | Even if they don't increase their GPU production capacity,
         | that's not "limiting" supply. It's keeping it the same. Only
         | now they can sell each unit for a larger profit margin.
        
       | jpecar wrote:
       | Servers? I thought they left even racks behind, they're now
       | selling these "AI factories".
        
       | czbond wrote:
       | Didn't they watch Silicon Valley to learn that lesson? Don't sell
       | the box.
        
       | thesuperbigfrog wrote:
       | What software will those Nvidia servers run?
       | 
       | Are they creating their own software stack or working with one or
       | more partners?
        
         | kj4ips wrote:
         | They have a Ubuntu derivative called DGX OS, that they use on
         | their current lines.
        
           | overfeed wrote:
           | I wonder which [publicly listed] companies would look at the
           | abandonment of Jetson and still commit to having Nvidia set
           | the depreciation schedule for them.
        
         | nijave wrote:
         | They already do have a pretty robust software stack that goes
         | all the way to code/analytics libraries. I'm not sure on the
         | current state of things but ~2020 they were automatically
         | testing chip designs for performance regressions in analytics
         | libraries across the entire stack from hardware to each piece
         | of software
        
       | kj4ips wrote:
       | It's my opinion that nvidia does good engineering at the
       | nanometer scale, but it gets worse the larger it gets. They do a
       | worse job at integrating the same aspeed BMC that (almost)
       | everyone uses than SuperMicro does, and the version of Aptio they
       | tend to ship has almost nothing available in setup. With the
       | price of a DGX, I expect far better. (Insert obligatory bezel
       | grumble here)
        
       | ecshafer wrote:
       | Nvidia already sells servers?
       | 
       | What I don't really get, is that Nvidia is worth like $4.5T on
       | $130B revenue. If they want to sell servers, why don't they just
       | buy Dell or HP? If they want CPUs want not buy AMD, Qualcomm,
       | Broadcom or TI? (I know they got blocked on their ARM attempt
       | before the AI boom) Their revenue is too low to support their
       | value, shouldn't they use this massive value to just buy up
       | companies to up their revenue?
        
         | jack_tripper wrote:
         | _> why don't they just buy Dell or HP?_
         | 
         | Why buy a complex but relatively low margin business that comes
         | with a lot of baggage they don't need, when they can focus on
         | what they do best and let Dell and HP compete against each
         | other for Nvidia's benefit?
         | 
         | Same reason why Apple doesn't buy Foxconn or TSMC.
        
         | mr_toad wrote:
         | > If they want to sell servers, why don't they just buy Dell or
         | HP?
         | 
         | They want to sell HPC servers, not general purpose servers.
        
         | andreasmetsala wrote:
         | Sometimes building a new organization is easier than trying to
         | improve a legacy one.
        
         | btian wrote:
         | But nVidia already sells servers (NVL72), and CPUs (Grace). Why
         | buy a bunch of overlapping companies?
         | 
         | And no sane regulator on the planet will allow them to takeover
         | AMD, Qualcomm, or Broadcom.
        
           | cyanydeez wrote:
           | Lucky for them Americans have recently gone insane.
        
       | thefourthchime wrote:
       | In a sense, they already do, since they're heavily invested in
       | CoreWeave. For those unfamiliar, CoreWeave was a crypto company
       | that pivoted to building out data centers.
        
         | zerosizedweasle wrote:
         | It's interesting to se the market try to do anything to rally.
         | The problem is you guys are rallying on the thought that you've
         | scared the Fed into cutting rates, but actually by rallying you
         | short circuit it. You ensure they won't cut. And that's how the
         | market's lillypad hopping thinking is actually just stupidity.
         | You rallied, so now there are no rate cuts so the crash will be
         | even more brutal.
        
         | wmf wrote:
         | GPU "neoclouds" are a different topic than whose logo is on the
         | server.
        
       | 2OEH8eoCRo0 wrote:
       | Why would they chase a lower margin business area? Are they out
       | of ideas?
        
         | nijave wrote:
         | More vertical integration
        
           | 2OEH8eoCRo0 wrote:
           | Like IBM?
        
             | wmf wrote:
             | Like IBM in the 1960s.
        
       | alecco wrote:
       | Guys, please read the article. Yes NVIDIA sells servers already.
       | What they mean is they are going to also do other system parts
       | that currenlty the partners are doing.
       | 
       | > Starting with the VR200 platform, Nvidia is reportedly
       | preparing to take over production of fully built L10 compute
       | trays with a pre-installed Vera CPU, Rubin GPUs, and a cooling
       | system instead of allowing hyperscalers and ODM partners to build
       | their own motherboards and cooling solutions. This would not be
       | the first time the company has supplied its partners with a
       | partially integrated server sub-assembly: it did so with its
       | GB200 platform when it supplied the whole Bianca board with key
       | components pre-installed. However, at the time, this could be
       | considered as L7 - L8 integration, whereas now the company is
       | reportedly considering going all the way to L10, selling the
       | whole tray assembly -- including accelerators, CPU, memory, NICs,
       | power-delivery hardware, midplane interfaces, and liquid-cooling
       | cold plates -- as a pre-built, tested module.
        
       | JCM9 wrote:
       | This would basically start to turn cloud providers into CoLo
       | facilities that just host these servers.
       | 
       | Makes sense longer term for NVidia to build this but adds to the
       | bear case for AWS et al long term on AI infrastructure.
        
       | nijave wrote:
       | I talked to someone at Nvidia ~2019 or ~2020 and their plan at
       | the time was to completely vertically integrate and sell compute
       | as a service via their own managed data centers with their own
       | software, drivers, firmware, and hardware so this seems like just
       | another incremental step in that direction.
        
         | SV_BubbleTime wrote:
         | In 2019 or 2020 that probably seemed reasonable.
         | 
         | Now? You would have to tell me nVidia was also building
         | multiple nuclear power plants to get the scale to make sense.
        
           | Kye wrote:
           | They're invested in nuclear power. Pairing datacenters with
           | small modular reactors is at least on the minds of all the AI
           | companies.
        
             | pishpash wrote:
             | One day they will build AI out of radioactive source
             | material directly and skip the reactor. Maybe.
        
         | noir_lord wrote:
         | That's one way to arrive at an IBM Mainframe like model I
         | guess.
         | 
         | It'll work until you can buy comparable expansion cards for
         | open systems (if history is any guide).
        
           | mrbungie wrote:
           | Yep, tech is incredibly circular. Once Nvidia gets there is
           | highly probable that "disruptive" competition will apear due
           | to mere desire/pressure for more freedom and options (and
           | knowing NVDA, also costs).
        
             | cyanydeez wrote:
             | AMD is already knocking on the same door. If they had
             | focused more on drivers it'd be an equal comparison.
        
         | michaelt wrote:
         | Maybe in 2019, but I find it hard to believe nvidia looks at
         | google's TPU business model enviously these days.
        
           | skybrian wrote:
           | Are TPU's a bad business model?
        
             | AnthonyMouse wrote:
             | Customers are increasingly wary of building their business
             | on someone else's land because they've seen it happen too
             | many times that once you're locked in the price goes up, or
             | the company who is now your only viable supplier decides to
             | enter your own market.
             | 
             | And at least if anyone can buy the hardware you'll have
             | your own or have multiple competing providers you can lease
             | it from. If you can only lease it and only from one
             | company, who would want to touch that? It's like purposely
             | walking into a trap.
        
               | cyanydeez wrote:
               | This might be the source of the AI bubble burst, just
               | like the 2000 bubble. Eventually someone's gonna raise
               | the price to cover a bill and suddenly everyone looks at
               | actual revenue and power bills , calculates within 6
               | months they'll not get their MBA turnip squeezed bonus
               | and walk away
        
               | skybrian wrote:
               | Lock-in is a valid concern, but on the other hand, for
               | many apps, it seems like this can be fairly easily
               | mitigated? If you can swap in a different LLM, I don't
               | think it matters if it's running on Google TPU's or
               | NVidia?
               | 
               | Meanwhile, at the hardware level, TPU's provide some
               | competition for NVidia.
        
         | amelius wrote:
         | > vertically integrate
         | 
         | Sounds like they are going the Apple way. How long until we
         | have to pay 30% to get our apps in their AI-Store?
        
           | DevKoala wrote:
           | We already pay a 100% premium on AWS.
        
             | amelius wrote:
             | At least that premium is not directly proportional to our
             | revenue.
        
           | diamond559 wrote:
           | More like IBM
        
       | modeless wrote:
       | They're not stopping at servers. They want to sell datacenters.
        
       | heisenbit wrote:
       | Competing with your customers can be a risky strategy for a
       | platform provider. If the platform abandons the neutral stance
       | its customers will be a lot more open to alternatives.
        
       | m_ke wrote:
       | Soon OpenAI will make its own chips and Nvidia its own
       | foundational models
        
       | wmf wrote:
       | I always wondered why a bunch of different companies make
       | identical graphics cards then complain that it's a horrible
       | business and Nvidia is screwing them. I wondered even more
       | strongly when I saw a dozen flavors of the NVL72 rack. If the
       | rack is so complex and difficult to manufacture, why have N
       | companies do redundant work?
        
         | AnthonyMouse wrote:
         | Designing the board is a different business from designing the
         | chip. You're negotiating with the DRAM fabs instead of the
         | logic fabs, selling to a thousand retailers instead of a dozen
         | integrators, etc. And once you've done all that, you do it more
         | than once. ASRock isn't just making Nvidia GPUs, they're making
         | AMD and Intel GPUs, motherboards, wireless routers, etc.
         | 
         | It's more efficient to have companies that specialize in making
         | all kinds of boards than to make each of the companies making
         | chips have to do that too. And it's a competitive market so the
         | margins are low and the chip makers have little incentive to
         | enter it when they can just have someone else do it for little
         | money.
        
       | foruhar wrote:
       | This video shows the systems being built and shipped with
       | cooling, cabling, etc.
       | 
       | It's pretty mind blowing what this crisis shows from the
       | manipulation of atoms and electrons all the way up to these
       | clusters. Particularly mind blowing for me who has cable
       | management issues with a ten port router.
       | 
       | https://youtu.be/1la6fMl7xNA?si=eWTVHeGThNgFKMVG
        
         | dzonga wrote:
         | what's mind blowing about the video you shared was the amount
         | of coper cable used.
         | 
         | I thought with fiber we wouldn't need coper cables maybe just
         | for electricity distribution but clearly I was wrong.
         | 
         | thanks for sharing
        
       | btown wrote:
       | "Nobody gets fired for choosing NVIDIA."
        
       | alberth wrote:
       | How does their attempt to acquire ARM (and failed) impact this?
        
         | wmf wrote:
         | It doesn't.
        
       | idatum wrote:
       | Can anyone comment on wafer-scale systems, multiple equivalent
       | chips on an entire wafer?
       | 
       | Seems like where things are heading?
        
         | wmf wrote:
         | Only Cerebras is doing wafer-scale. It seems to be working for
         | them but no one is copying them. The minimum unit (one wafer)
         | costs millions and it's not clear how good their multi-wafer
         | scaling is.
        
       ___________________________________________________________________
       (page generated 2025-11-14 23:01 UTC)