OpenAI models are coming to Google Cloud | Hacker News Reader

OpenAI models are coming to Google Cloud

openai.com

326 pointsarmcat152 comments

Sergey better really go into "founder mode" or else Thomas Kurian (head of GCP) will continue to eat Deepmind's lunch.

Between selling TPUs to external companies and now hosting competitor models, Gemini just doesnt have the full faith of the Alphabet capital allocators, and I can't really blame them. A mature company like google would rather just expand their extremely profitable cloud business rather than dump billions into a black hole on the off chance they achieve RSI or superintelligence or whatever.

Google uses Gemini for its AI Overview on Google Search results, though. And it seems to me that this is an area where they really have a big incentive to run their own AI that they completely control. AI Overview sits at the top of the Google Search results first page, so when it comes to marketing it's sitting in the most coveted place on the entire Internet.

It's also marketing that their AI is hilariously terrible because AI Overview is the biggest posterchild for AI slop and often used as an example of how bad Google is at AI. Gemini proper isn't that bad, just the ultra cost cutting version on search is.

To be fair the search version of Gemini is nowhere near as bad as it used to be. The bar is in the floor I know.

> The bar is in the floor I know.

People don't seem to realize, this is how successful businesses happen. Go take a look at Windows 1.0, say.

It has gotten a lot better and it's useful for information on most normal search queries now.

Yes, I'm realistic about it too. For simple queries it nails it more often than not. It's only asking it more complex technology related questions that it falls apart for me, and in those cases I already know I should be talking to one of the big boi models for certain lines of questioning.

Google certainly ain't gonna run a several T model for every search that's for sure (well, not yet).

Do you think most people around the world want access to ultrathink?

For a lot of folks, something low cost and low intelligence is what they want. And for a few that needs high intelligence, there's stronger models.

Google has both bases covered. It is brilliant.

I've had the Search version of Gemini give substantially better responses than (free) ChatGPT in some cases.

im talking about research compute allocation

It’s not called Gemini Overview…

What does “capital allocators” mean here? If your implication is that the Gemini org is resource starved and a few extra dollars will put their model in the lead that is 100% incorrect. If anything it’s the bloat and bureaucracy keeping it behind.

Signing partnerships and hosting third party models costs Google nothing, and in fact brings in a ton of revenue that will eventually go back into Gemini.

Infra might very well be the right play. Over the past year, it seems like the evidence is on the side that there's no "secret sauce" in AI, and models will all converge towards similar capabilities. Not much margin to be had for having the best model for a month or less until another lab leapfrogs you by 3%.

I think RAM turned out to be the secret sauce all along!

I keep saying LLMs are the new VMs and everyone gets mad at me, but time will tell.

If there's no point in having the best model, there's also no point in leapfrogging the current best model. Conversely, labs trying to leapfrog each other indicates that they consider the advantages of having the best model (or at least a better one than they had before) worth it, even taking the competition into account.

Suppose the big 2 labs didn't think the expense was "worth it."

Wouldn't they keep spending anyway because they need to hype things up before IPO?

I think it shows that they consider their to be advantages to being seen as having the best model, which is a subtly different thing.

"4 out of 5 dentists prefer DentifreshAI"

I mean the labs are in a tricky spot where they have to justify massive valuations. "We're going to stay where we are and focus on inference" isn't the growth story you need when you're trying to justify a 2T valuation on 80B in revenue (unless you dress it up as a safety concern).

Google can afford to take their foot off the gas without killing their business.

> Our cloud business is at any given moment in time forgoing short-term revenues because we are making compute trade-offs to support frontier AI development

Sundar Pichai just this week.

https://finance.yahoo.com/technology/ai/articles/google-ceo-...

This can be just as well phrased as:

> Our frontier AI business is at any given moment in time forgoing short-term development because we are making compute trade-offs to support our cloud business revenues.

Exactly, the definition of tradeoffs

Infact all companies, at all times are doing tradeoffs. You and I can be doing something else right now but I'm here on HN replying to you - and that's a tradeoff.

If someone suggests -just do X then you don't have to do a tradeoff, that just sounds wrong to me.

The primary trade off amounts to CEOs enriching themselves by sucking resources out of the company they control. They trade the company's future and their employee's incomes for personal wealth. Investors are the bag holders for CEOs and their insiders.

Google is a multi trillion dollar company. Sundar Pichai is only recently a billionaire.

The investors are the ones whose net worth is measured in trillions, and most of the individuals are invested through 401k and pension accounts.

We’ve been seeing compute allocation failures in GCP us-central1 for about a month now- that can’t be related can it? These aren’t TPUs, this is normal compute for dataflow/etc jobs.

Frontier AI development via agents uses plenty of CPU compute. Maybe in a lopsided way during development as opposed to training?

10 bucks say it's just a high-demand region and GCP's hardware supply chains are at capacity. So you see stockouts.

Google are now offering Claude models in the Gemini "app". https://cloud.google.com/blog/products/ai-machine-learning/w...

I think it makes sense for them to separate the application/productivity layer from the model layer. And offer models from many model providers, including OpenAI.

To stay competitive with Microsoft and other productivity suites.

AI hosting currently generates revenue, so the choice is trivial. Building an LLM model is company PR, while hosting another’s model provides immediate income.

The tragic truth is that we live in a world where open models are perpetually only a few months behind frontier models.

Opus 4.6 was "good enough." Having a frontier model is no longer a requirement to print cash with a GPU.

I don't understand all the wild speculation and question here - openai is selling a shit ton of inference through third parties. these third parties want to provide inference to their customers who use gcp (and don't want to also inference through bedrock or azure ai foundary.

these are all zdr engagements and thus "enterprise" contracts.

basically it boils down to - anthropic already provides inference direct through GCP, Azure, and AWS - so to compete openai is doing the same.

Preparing for Oracle bankruptcy.

There's always a silver lining!

Do not anthropomorphize Larry Ellison.

I could only hope...

Does anyone still remember OpenAI’s founding? I had thought that OpenAI didn’t want to give the weights to Google because the original reason for its existence is to prevent Google from achieving AGI.

> Enterprises with Google Cloud spend commitments will be able to draw down their GCP commitments for these purchases.

This is a fairly big move by Google... It means revenue they have already booked and reported to shareholders they now might have to give away - potentially all of it. Bean counters hate that!

“Reported to shareholders” doesn’t really mean anything. There are thousands of such line items on financial reports. The ones that matter, such as recognized revenue, don’t count future bookings. It is understood that customers renege on these commitments all the time.

Google is simply telling customers - there’s a new Google Cloud service, and you can use it under your existing account.

Let's not quite right... you don't recognize revenue in a consumption business until the usage actually happened

pretty sure that's what the "spend commitment" changes. It's a "use it or lose it" agreement. It's as good as a debt the moment it is signed, used or not.

I think they declare it as pipeline revenue or something instead of current year revenue

RPO remaining performance obligations (not on balance sheet) unless prepaid then its deferred revenue (balance sheet). When it’s consumed it is recognized as revenue.

GAAP does not permit booking contracted future revenue onto the balance sheet or income statement, because the spend is not actually an unconditional debt until the service has been rendered.

It doesn't matter if it's committed or not, it still won't be recognized on the formal financial statements.

My understanding is that "use it or lose it" means you book it the last day of the contract or similar. Anyway not at the front so same point applies

That’s not the way it works. Google still recognizes the revenue. The cost of goods sold and gross profit are what get hit here but it’s no different than any other partner than can sell their product through the Google marketplace and take advantage of committed spend.

No. It means customers can turn booked revenues into billed revenue by consuming OAI models on GCP (pass-through contracting). OAI will get their money, too, but Google will charge the customer for actual platform consumption. This was previously missing and GCP customers who wanted ChatGPT or Codex were limited to just making API calls from their GCP projects .

We have asked them about this, openai spend only counts as partnership and only up to 25% of your commitment can be spent on partners

Google doesn't have enough compute to serve there own customers.

I have AI pro via a promo offer with my phone and Antigravity has been almost unusable the past week, exceptionaly slow and they completely nerfed the promo accounts by not giving access to Argon or the new claude models as well as making Flash 3.8 much worse.

Complete bait and switch from the company that advertises that the phone comes with $227 worth of a subscription, then later turns on a flag that marks these accounts as promo and cut access to it because of it. But i understand why, because they simply don't have compute to handle it, but they are still giving out these accounts.

Gotta love the disclaimer at the bottom:

  Microsoft remains OpenAI’s primary cloud partner.

Things that you only say because you are contractually forced to say them.

Or don't want to crash stock price.

OpenAI buys from Azure because of the Microsoft investment deal.

OpenAI sells and reaches enterprise customers via GCP.

Enterprises want zero data retention guarantees and they're already on AWS or GCP, so this is how OpenAI reaches those customers and makes more money.

isn't the largest portion of openAI's reach to (overwhelming cloud provider leader AWS) enterprise customers via amazon bedrock?

OpenAI only launched on bedrock ~3 months ago, so I doubt it. It has been running on azure since day 1.

so.. yeah i guess bedrock is their largest distribution system.

Microsoft has had ZDR for inference since pretty much the beginning. Just needed to fill out a form and get approved.

However, consider this:

> Enterprises with Google Cloud spend commitments will be able to draw down their GCP commitments for these purchases.

What does this mean?

You can use your preexisting committed spends for GCP on Open AI, i.e. if an enterprise 10M/year contracted spends for first party GCP services, now they can also use that on OAI models hosted on GCP.

Meaning they can swap out compute/storage spends on GCP to LLM model spend or use left over budget ( which would go waste anyway ) and so on . Any OpenAI sales by GCP teams would be considered against their quota typically in this type of arrangement [2]

This is normally not done to third party[1] marketplace services, Azure already does this for OpenAI, it helps drive some enterprise spends as it is from for fungible budgets or different approving authority internally in a org when it is not a new vendor, new budget etc.

----

[1] It is not too hard to get into these types of programs i.e. it not only for OpenAI level partners mid sized startups can also be part of these deals in marketplace, the requirement usually is minimum size and the partner having to prove/show that are spending that customer's revenue on the compute etc so money is going back to them, then AWS/Azure/GCP will consider an org for this type of program.

Given that GCP is hosting these models on their cloud and therefore OpenAI is spending a % of every dollar back on GCP for every token sold like this; it wouldn't have been too hard for OpenAI to got such a program even if they were a mid sized startup.

[2] Multiples maybe involved (also used to push some first party products) 1x would 1M in OAI sales counts as 1M in their Qtr/annual targets 0.5x would mean 500k and 5x means 5M quota would be retired for 1M in sales etc.

So when cloud deals are signed, it is with a spending figure? Do you get discounts if it is a large number?

Thanks for you reply, it was very enlightening

Yes; larger discounts with more commits, you also get some price guarantees on SKUs quite valuable these days and other benefits - like say premium support, architecture guidance, access to product teams and so forth.

Self-service retail buyers of cloud have access to some of these price advantages as well, 1-3 Year reservations(even paid monthly) are 40% cheaper of the bat after all. PAYG with fluctuating usage is quite more expensive than committed consumption for cloud.

The way they work, you get a preferred price for various services at a substantial discount to list rate, but you commit to a guaranteed spend level for that period. The guaranteed spend is typically account-wide, but you have to pay a true-up at the end of the period if you fail to hit it.

It's a bit of calculus: will your growth increase enough such that, especially with the discount, you'll reach the minimum?

You can see these agreements littering 10-Qs, or similar, like when Snapchat filed their S-1 and revealed:

>When Snap originally filed for its IPO in 2017, it revealed a massive $2 billion commitment over five years ($400 million annually) with Google Cloud.

Is it really gcp hosting openai in their cloud? Afaik is a third party marketplace so it looks like you're buying openai saas via gcp but not hosted in gcp. That's key difference with azure specially for egress costs if your data leaves gcp to openai right?

From the press release this thread is about (emphasis mine) -

> backed by Google Cloud’s data governance and high-performance infrastructure

Yes but does it's not clear for me. It could be just a private interconnect to openai infrastructure. My guess is that is not hosted in gcp... That's why I'm asking

Only a GCP rep can give you an authoritative answer; but in my experience hyper scalers don’t promote services that not hosted on their platform .

Enterprise* customers are overwhelmingly more likely to primarily host on Azure than GCP, m365 runs the corporate world top to bottom. OpenAI doesn't care where the customers come from, more platforms == more reach, they have zero reason to be preferential to any single hosting platform.

* By "enterprise" here I mean "big company that is not a tech company". If most of your userbase is carrying a mac, you're probably a tech company. If you're mostly making widgets in factories, chances are high Windows and the ecosystem around it is involved.

I've heard this described as having a "real" or "fake" corporate job on social media.

If you have a Thinkpad, you have a real job. And if you have a MacBook, you have a fake job.

What if I have a MacBook which is primarily used for ssh'ing into Linux machines?

You might be Chinese!

Fake. We all know you actually just talk to an agent all day.

What if I have a windows laptop but the only windows apps used are teams and Outlook, everything else runs in WSL? :-)

Running m365 doesn't meant their boring ERP system has to run on Azure even if MS would love them to.

These are the companies running SAP and Oracle. It can run anywhere, including GCP and AWS.

Isn't Google renting cloud compute from XAI?

Yep, and they pay them like 900 mil per mo for it, pretty crazy from the company that first brought TPUs to market

Google also has a stake in xAI... so they are financing themselves too

AI centipede

string theory Ouroburos

That’s putting it lightly. Google was one of the biggest early investors in SpaceX and is currently the second largest shareholder after Musk.

Would have been more accurate to write “We are contractually obliged to say that ‘Microsoft remains our primary cloud provider,’ whatever that means.”

and 49% of OpenAI profits

Google should just spin out DeepMind as an independent company and get out of the model building business: just sell tents and shovels as they say, and leave the actual (gold) mining to others.

Google Search is still a large fraction of Alphabet's revenue, and Alphabet needs to be able to deliver AI results for people's searches if it wants to stay relevant in the search game.

I imagine it's much easier to have fine grained control over AI search results if you develop and own the AI yourself.

Last I checked their AI results were terrible, worse than anything I get from ChatGPT or Claude. The only benefit is their AI results are fast.

If I were google there's no way I'd leave their AI overviews in place as-is.

I was thinking about them fine-tuning the results for marketing purposes. But I just did a bit of research and it seems that would probably actually be illegal. So never mind, maybe?

Can you give an example of a search query with a terrible AI Overview? Or even better, a class of search queries that produce terrible AI Overviews? Would really appreciate it!

Sure, I'll try. It has been a while (in LLM terms) since I opened them.

Any news?

Search for: "hanson scary movie 2"

AI overview result: images of Chris Elliot as Hanson and an AI summary of the band Hanson.

https://imgur.com/a/IVhuAUE

This is an extension of the long-standing issue where Google just ignores keywords randomly, even when you only typed a few (it'd be excusable if it dropped one in ten, say, and zero out of 3/4.)

Thank you!

it used to be, i'm finding myself following up the overview conversation more often than not

My work experience with trying to access AI models from GCP has been rather spotty. Gemini works great for work we need it to do, but dynamic usage limits are hitting us hard. And while we tried to use Anthropic models on there, we could not get them fully enabled, for one reason or another.

Now two companies get to steal your work instead of one.

Listen to people like Buckmaster who admits he will never put anything meaningful into codex again.

If you are an academic, figure out why your university is not running local AI models in house for academics and students to use. It's absurd that any college would want their entire student body datamined by a corporation who will use that data to destroy job prospects so a billionaire CEO can become a trillionaire.

Colleges should be running caches you can set your browser to use when doing research so all the pages are stored. All the info should be used to improve in house llms. It is no different than when universities funded super computers, computing labs, web host space, academic journals/resources, and everything else. It is like all universities stopped being universities. They charge students more than ever and provide none of the resources they used to.

How are they trusting Google not to destill them

Probably a very strict contract, and letting OpenAI read all the session logs like they do for their own customers.

Nvidia GPUs let you load encrypted weights onto them and to decrypt on chip so Google would need to have some sort of Nvidia management firmware attack to read from HBM while the GPU is live

Doesn’t really matter what happens at the GPU level when Google can see all API traffic.

Except google cloud doesn't use Nvidia GPUs but TPUs.

Google has both. Afaik we don't know whether OpenAI models will run on TPUs or GPUs here.

I bet it would be the NVIDIA, at least initially, not for the encryption capability per se, but for simple fact that adapting to TPU at high performance is non-trivial.

Isn't distillation more of a black box type affair?

Distillation does not require access to weights.

Google Cloud couldn't give a shit about Google Deepmind, and backdooring Clouds customers only stands to hurt Cloud.

This is speculation, but funnily enough Gemini models show the least evidence of being distilled from OAI or Anthropic compared to everything else in the market.

I feel this is correct. Google has probably the most data of any company in the world. For better or worse, they don't need to train on the outputs of other models.

Contracts and lawyers are more powerful than any technical tool.

Yeah like the copyright lawyers of content owners that got their works trained on?

How many trillions of dollars were those content owners worth?

Enough. They just weren’t working in coordination.

You are vastly overestimating the power of publishers. The top book publisher in the US is worth $5 billion. All publishers together are about $50 billion. The New York Times is $11 billion. Fox is about $15-25 billion. Add together every single book publisher, newspaper, magazine, TV/radio, digital publisher and really every print and broadcasting platform in the country and you’d reach about $100-$150 billion. That is roughly 3% of Google.

These companies have a ton of cultural capital, sure, but in American capitalism that means nothing.

GCP already hosts Anthropic models so probably trusts and lawyers.

America is a high-trust society.

More like - America is a high-litigation society.

I don't believe threat of litigation alone explains the extent of it.

The way an American company, esp a reputable one like Google, conducts itself internally is different. There will be much more internal backlash to do the wrong thing way before litigation.

That's Hollywood narrative for you. Lawyers want cases with damaged plaintiffs and work on contingency.

Japan is a high-trust society. America is the opposite.

Everything is relative, but fair enough.

The US is medium trust.

First answer this: How are they trusting OpenRouter not to distill them?

Open router doesn't host the models, so they wouldn't have access to the weights or reasoning traces right?

They can still collect (input,output) pairs.

OpenRouter doesn't see the encrypted reasoning traces.

But they do see input and output, which gives them a treasure of training data.

At a org like Google Cloud barring rogue L5/6/7s acting on their own any sniffing of distillation will get leaked. Too many chaos exist at a org size of Google to systematically do it given the contract at hand. This is a company where every spicy memo or mild gossip will risk leaks.

I'd bet the distillation balance-of-trade goes the other way on that one

And Google is renting compute from SpaceX.

Google is also the second largest shareholder of SpaceX.

We need a new acronym

Google, NVIDIA, OpenAI, MSFT…

GNOM…oh wait no

The strategy that I’ve heard said out loud now by all the major technology companies is “we’re all doing it (AI) together”

In my view this was more or less the political result of the White House summit

So I think we’re seeing what is basically going to be a giant amorphous set of collaborations such that nobody gets taken down, and the whole market doesn’t collapse

If Alphabet, OAI, MSFT, AMZ…etc all agree to be interoperable then they will be acting more or less as a single body in tension.

Whether that’s good or not is a question of your worldview but I think that’s the strategy

This is like that SNL skit with Tim Meadows where Will Ferrell has the F on his shirt on the kids TV show and keeps getting in the inappropriate order

Add in Meta (Muse app is a winner)

MOM-NG GNOMM NGMOM MGNOM MGMNO MGMON ONMGM NOMGM NOMMG

Giant companies working together isn’t an AI-age phenomenon. Go back 10, 20, 30 years and you’ll see that despite the public perception of competitiveness all the largest tech companies (Apple, Microsoft, Google, Samsung, Amazon) buy/sell to each other, sign partnerships and have tight economic interdependencies.

Who said it’s new? :)

However if you look at the trust busting era what preceded it was this exact kind of private “cooperation” to prevent government involvement

Add Anthropic and you get MANGO. MSFT, Anthropic, NVIDIA, Google, OpenAI.

Let’s go with Mango!

Does this mean OpenAI is going to use TPU now?

وبن حمارا

Test

Is this the begging of the end of Gemini?

No? Why would you think that’s the case? Gemini fills its own niche

Gemini and GPT fill the same niche. They are direct competitors.

They overlap but are positioned differently, Gemini fills a different niche. In any case it would be a pretty dumb idea to ditch your own models, optimized for your own hardware, just because you’re also offering a competitor. Google can learn usage patterns for GPT models while continuing to work on their own models. No reasons to believe they will just stop developing Gemini

GCP has been hosting Anthropic models for a while.

Google owns Anthropic stock.

Why would it? The point of Gemini is provide user facing services cheaper than if they were paying OpenAI, make some money on the side with API credits and position themselves as a leader when investors want to see AI everywhere.

Google Cloud is competing with AWS. Amazon has Bedrock so Google needs to keep up with them by serving both Claude and GPT, along with Gemini and open models.

"OpenAI models will soon be available in the Gemini Agent"

They definitely need to have a better name if they're going to use 3p models

New to Google?

Prediction: google cloud will pivot to being a hyperscaler versus a model + hyperscaler. Kind of like how xAI started hosting anthropic workloads with colossus cluster.

bc the alpha is not the model, the alpha is having the physical infra (chips, power, etc).

But x.ai also hosts Grok? Why would google stop developing their own models? I think you're right that Google will try to make money from their infrastructure (that's basically what Google Cloud is too).

As much as I like Elon, if Grok was a hero shooter its concurrent steam player count will be in low hundreds.

Is that a long term prediction? Are you saying that physical infra will always lag demand from here on?

A company is allowed to do two things

>Kind of like how xAI started hosting anthropic workloads with colossus cluster.

Google has been hosting part of Anthropic's load long before that deal. They had a big degradation due to compiler bug for TPUs in an optimization they made to work around TPUs not being able sample from the logits efficiently.

Reply on news.ycombinator.com