I have seen commentary floating around that the Trump administration is, to some extent, looking to build the Arab world into stronger allies for the US. The implication being that there was a possibility that they might fall into China's sphere of influence.
I could see a 'two birds one stone' motive where in doing this the US not only gets to 'friendshore' some of its capacity without having to deal with NIMBYs, but also strengthens the alliance and keeps the Arab power players firmly in the US camp.
Finally, the U.S. government is promoting a U.S.-led global technology ecosystem within which AGI can be pursued. For example, the U.S. government recently supported Microsoft’s expansion into the United Arab Emirates to develop new data centers, in part to prevent Chinese companies from entrenching their position.
It always seemed outlandish that in The Animatrix, the first AI city (01) was located in the Middle East...
If we had limitless time, it would be interesting to know how this happened. I guess the prehistory of it involved Saudi Vision 2030 (e.g. the desert city Noem), and the general hypermodernization of Dubai. You can see precursors in the robot Sophia getting Saudi citizenship in 2017, and the UAE's "Falcon" LLM in 2023.
But the initiative must have come from the American side - some intersection of the geopolitical brain trust around Trump, and the AI/crypto brain trust around David Sacks. The audacity of CEOs who can pick a country on the map and say, let's build a whole new technological facility here, combined with the audacity of grand strategists who can pick a region of the world and say, let's do a huge techno-economic deal with our allies here.
There must have been some individual who first thought, hey, the Gulf Arabs have lots of money and electricity, that would be a good place to build all these AI data centers we're going to need; I wonder who it was. Maybe Ivanka Trump was telling Jared Kushner about Aschenbrenner's "Situational Awareness", they put 2 and 2 together, and it became part of the strategic mix along with "Trump Gaza number one", the new Syria, and whatever they're negotiating with Iran.
Sam Altman was publicly talking about this in 2024-02 (WSJ). I think this was the 1st time I've encountered the idea. Situational awaness I think was published ~4 months later, 2024-06 (https://situational-awareness.ai/ says "June 2024")
Other countries adopting Chinese hardware may mean this was basically the US being forced to sell their GPUs to prevent the Chinese from taking advantage of economies of scale.
Seriously, is there nowhere in America we can make this happen at scale? If we wanted to, we could do this ourselves easily. We have the natural gas, even if nuclear would be too slow to come online.
It's only 5 GW, and the US average is ~440 GW. The US would not have to build any more power plants - just run the ones it has more. It could just reduce liquefied natural gas exports and produce another >25 GW electrical average.
Thanks for writing this up. I spent 2-3 days this month encouraging the UAE to set up an AI Security Institute. I was in Abu Dhabi during President Trump's visit and spoke briefly with a few government officials whose title was "Chief AI Officer" of ministry such-and-such. I mostly wanted to know if they'd be interested in potentially collaborating with other AISIs. However, I'm not planning to continue working on UAE AISI stuff in the future.
Our government, having withdrawn the new diffusion rules, has now announced an agreement to sell massive numbers of highly advanced AI chips to UAE and Saudi Arabia (KSA). This post analyzes that deal and that decision.
It is possible, given sufficiently strong agreement details (which are not yet public and may not be finalized) and private unvoiced considerations, that this deal contains sufficient safeguards and justifications that, absent ability to fix other American policy failures, this decision is superior to the available alternatives. Perhaps these are good deals, with sufficiently strong security arrangements that will actually stick.
Perhaps UAE and KSA are more important markets and general partners than we realize, and the rest of the world really is unable to deploy capital and electrical power the way they can and there is nothing we can do to change this, and perhaps they have other points of strategic importance, so we have to deal with them. Perhaps they are reliable American allies going forward who wouldn’t use this as leverage, for reasons I do not understand. There are potential worlds where this makes sense.
Diplomacy must often be done in private. We should not judge so quickly.
The fact remains that the case being made for this deal, in public, actively makes the situation seem worse. David Sacks in particular is doubling down and extending the rhetoric I pushed back against last week, when I targeted Obvious Nonsense in AI diffusion discourse. Even within the White House, the China hawks are questioning this deal, and Sacks responded by claiming to not even understand their objections and to all but accuse such people of being traitorous decels wearing trench coats.
I stand by my statements last week that even if accept the premise that all we need care about are ‘America wins the AI race’ and how we must ‘beat China,’ our government’s policies, on diffusion and elsewhere, seem determined to lose an AI race against China.
This is all on top of the entire discussion not only dismissing but outright ignoring the very real possibility that if anyone builds superintelligence, everyone dies. Or that everyone might collectively lose control over the future, with other bad outcomes. Once again, in this post, I will do my best to set these concerns aside.
This ‘have to beat China’ hyperfocus out of Washington has reached new heights of absurdity. I offer an off topic example to drive the point home before we dive into AI.
I am going to go ahead and say, I want us to beat China, but if China cured cancer then that would be a good thing. And indeed it would reduce, not increase, the urgency of America needing to cure cancer.
If I join the war on cancer, it will not be on the side of cancer.
The Central Points From Last Week
The point of the diffusion rules is to keep the AI chips secure and out of Chinese hands, both in terms of physical security and use of their compute via remote access. It is possible that the agreements we are making with UAE and KSA will replace and improve upon the functionality, in those countries in particular, of the diffusion rules.
It’s not about a particular set of rules. It is about the effect of those rules. Give me a better way to get the same effect, and I’m happy to take it. When I say ‘something similar’ in #2 and #4 below, I mean in the sense of sufficient safeguards against the diversion of either the physical AI chips or the compute from the AI chips. Access to those chips is what matters most. Whereas market share in selling AI chips is not something I am inclined to worry about except in my role as Nvidia shareholder.
I would also clarify that in #3, I definitely stand by that I do not consider them reliable allies going forward, and there are various reasons that even the best version of these agreements would make me deeply uncomfortable, but it is possible to reach an agreement that physically locates many data centers in the Middle East and lets them reap the financial benefits of their investments and have compute available for local use, but does not in the most meaningful senses ‘hand them’ the compute in question. As in, no I do not trust them, but we could find a way that we do not have to, if they were fully open to whatever it took to make that happen.
If you told me I was wrong about something here, my guess would be that I was wrong about the geopolitical situation, and UAE/KSA are more important strategic partners or more reliable allies than I realize. World geopolitics is not my specialty, and I have uncertainty about these questions, which of course runs in both directions. Discussions in the past week have updated me a small amount in the direction that they are likely more strategically important than I realized.
I also would highlight the implicit claim I made here, that the pool of American advanced AI chips is essentially fixed, and that we have sufficient funding available in Big Tech to buy all of them indefinitely. If that is not true, then the UAE/KSA money matters a lot more. Then there is the similar question of whether we were going to actually run out of available electrical power with no way to get around that. A lot of the question comes down to: What would have counterfactually happened to those chips? Would we have been unable to deploy them?
With that in mind, here are the central points I highlighted last week:
America is ahead of China in AI.
Diffusion rules serve to protect America’s technological lead where it matters.
UAE, Qatar and Saudi Arabia are not reliable American allies, nor are they important markets for our technology. We should not be handing them large shares of the world’s most valuable resource, compute.
The exact diffusion rule is gone but something similar must take its place, to do otherwise would be how America ‘loses the AI race.’
Not having any meaningful regulations at all on AI, or ‘building machines that are smarter and more capable than humans,’ is not a good idea, nor would it mean America would ‘lose the AI race.’
AI is currently virtually unregulated as a distinct entity, so ‘repeal 10 regulations for every one you add’ is to not regulate at all building machines that are soon likely to be smarter and more capable than humans, or anything else either.
‘Winning the AI race’ is about racing to superintelligence. It is not about who gets to build the GPU. The reason to ‘win’ the ‘race’ is not market share in selling big tech solutions. It is especially not about who gets to sell others the AI chips.
If we care about American dominance in global markets, including tech markets, stop talking about how what we need to do is not regulate AI, and start talking about the things that will actually help us, or at least stop doing the things that actively hurt us and could actually make us lose.
Diffusion Controls Have Proven Vital
Diffusion controls on AI chips we’ve enforced on China so far have had a huge impact. DeepSeek put out a highly impressive AI model, but by their own statements they were severely handicapped by lack of compute. Chinese adoption of AI is also greatly held back by lack of inference compute.
China is competing in spite of this severe disadvantage. It is vital that we hold their feet to the fire on this. China has an acute chip shortage, because it physically cannot make more AI chips, so any chips it would ship to a place like UAE or KSA would each be one less chip available in China.
Dean Ball (White House Strategic Advisor on AI): cue the @ohlennart laser eyes meme.
South China Morning Post: China’s lack of advanced chips hinders broad adoption of AI models: Tencent executive.
Washington’s latest chip export controls could widen the gap in AI adoption between China and the US, Tencent Cloud’s Wang Qui says.
Whenever you see arguments from David Sacks and others against AI diffusion rules, ask the question:
Is an argument for a different set of export controls and a different chip regime that still protects against China getting large quantities of advanced AI chips?
As in, that what matters is our market share of AI chips, not who uses them?
This is not a strawman, for example Ben Thompson argues exactly this very explicitly and repeatedly.
Indeed, Ben Thompson’s recent interview with Jensen Huang, CEO of Nvidia, made it clear both of them have this exact position. That to maintain America’s edge in AI, we need to sell our AI chips to whoever wants them, including China, because ‘China will not be held back’ as if having a lot more chips wouldn’t have helped them. And essentially saying that all Nvidia chips everywhere support the ‘American tech stack’ rather than China rather obviously turning around and using them for their own tech. He explicitly is yelling we need to ‘compete in China’ or else.
Complete Obvious Nonsense talking of his own book, which one must remind oneself is indeed his job, what were you really expecting him to say? Well, what he is saying is that the way we ‘lose the AI race’ is someone builds a CUDA alternative or steals Nvidia market share. That his market is what matters. It’s full text. Not remotely a strawman.
I would disagree with arguments of form #2 in the strongest possible terms. If it’s arguments of form #1, we can talk about it.
Lennart Heim: To put the new 5GW AI campus in Abu Dhabi (UAE) into perspective. It would support up to 2.5 million NVIDIA B200s.
That’s bigger than all other major AI infrastructure announcements we’ve seen so far.
In exchange for access to our chips, we get what are claimed to be strong protections against chip diversion, and promises of what I understand to be a total of $200 billion in investments by the UAE. That dollar figure is counting things like aluminum, petroleum, airplanes, Qualcomm and so on. It is unclear how much of that is new.
The part of the deal that matters is that a majority of the UAE investment in data centers has to happen here in America.
I notice that I am skeptical that all the huge numbers cited in the various investment ‘deals’ we keep making will end up as actual on-the-ground investments. As in:
Walter Bloomberg: UAE PRESIDENT SAYS UAE TO INVEST $1.4T IN U.S OVER NEXT 10 YEARS
At best there presumably is some creative accounting and political symbolism involved in such statements. Current UAE foreign-direct-investment stock in the USA is only $38 billion, their combined wealth funds only have $1.9 trillion total. We can at best treat $1.4 trillion as an aspiration, an upper bound scenario. If we get the $200 billion we should consider that a win, although if the deal is effectively ‘all your investments broadly are in the West and not in China’ then that would indeed be a substantial amount of funds.
Nor is this an isolated incident. The Administration is constantly harping huge numbers, claiming to have brought in $14 trillion in new investment, including $4 trillion from the recent trip to Arabia, or roughly half of America’s GDP.
Jason Furman (top economic advisor, Obama White House): That’s nuts and baseless. I doubt the press releases even add up to that. But, regardless, press releases are a terrible way to determine the investment or the impact of his policies on it.
Justin Wolfers: Trump has claimed a $1.2 trillion investment deal from Qatar. Qatar’s annual GDP is a bit less than $250 billion per year. So he’s claiming an investment that would require every dollar every Qatari earned over the next five years.
Not that the numbers ultimately matter all that much. What does matter is: How will we ensure the chips don’t fall literally or functionally into Chinese hands?
It comes down to the security provisions and who is going to effectively have access to and run all this compute. I don’t see here any laying out of the supposed tough security provisions.
Without going into details, if the agreements on both physical and digital security are indeed implemented in a way that is sufficiently tough and robust, if we are the ones who both physically and digitally control and monitor things on a level at least as high as domestically, and can actually have confidence none of this will get diverted, then that goes a long way.
We don’t yet have enough of that information to say.
The public explanations for the deal, and the public statements about what safety precautions are considered necessary, do not bring comfort.
I very much do not like comments like this, made in response to the Bloomberg piece above.
David Sacks (US AI Czar): If the concern [about the deal] is about diversion of advanced semiconductors to China, that’s an important policy objective but one that is easily addressed with a security agreement and a “trust but verify” approach. Modern AI supercomputers are 8 feet tall and weigh two tons; these are not like diamonds smuggled in a briefcase. To verify that chips are where they’re supposed to be, an auditor can just visit the data center and count server racks.
Semianalysis: As such, physical inspections are key to ensuring the GPUs are where they are supposed to be. The White House is aware of the risks and is prepared to ensure the chips are where they are supposed to be. As Bloomberg reported (quotes David Sacks’s statement above).
I’m sorry, what? You can ‘just visit the data center and count server racks’?
It terrifies me to see so many people arguing for the deal explicitly saying that this is a good statement, rather than a terrifying statement that we are hoping no one involved in the actual work believes to be true.
To be clear, I think diversion via remote access is far more likely than physical diversion, and this response does not address the remote access issue at all, but even simply treating this as a physical diversion issue, o3, is the quoted statement accurate?
o3 (full one shot, full response quoted, in response to literally quoting Sacks without attribution and asking ‘is this accurate?’):
Short version: Mostly wrong. Chips are tiny and high-value, so diversion is hard to stop with “count the racks.” Physical audits help but they are far from sufficient.
Why “just count server racks” fails
High value-to-weight ratio.
At ~$30 k per H100, a single 20 kg carry-on full of GPUs is > $1 m. That’s an attractive target. Racks are heavy; the payload isn’t.
Modularity.
GPUs slide out in minutes. An auditor sees the enclosure; they cannot tell if the trays were swapped yesterday or will be tomorrow. Dummy blanks cost <$10.
Front companies & trans-shipping.
Chips often leave the compliant data-center before any inspection window. Paperwork shows them installed; reality changes after the auditor flies home.
Remote lease risk.
Even if the hardware stays in the U.S., Chinese entities can simply rent time on it. Export-control rules treat “deemed exports” (remote access by restricted users) as violations.
Limited audit frequency & competence.
Meaningful verification requires random, intrusive, technically savvy inspections plus cryptographic attestation of firmware. Counting racks is neither.
Bottom-line probability estimate
My best estimate: ≤ 30 % that “security agreement + occasional rack counting” alone keeps advanced GPUs out of China for > 3 years. ≥ 70 % that significant leakage continues absent tighter controls (HW tracking, cryptographic attestation, and supply-chain tagging).
So the quoted claim is misleading: rack-level audits are helpful but nowhere near “easily addresses” the diversion problem.
When I asked how many chips would likely be diverted from a G42 data center if this was the security regime, o3’s 90% confidence interval was 5%-50%. Note that the G42 data center is 20% of the total compute here, so if we generously assume no physical diversion risk in the other 80%, that’s 1%-10% of all compute we deploy in the UAE.
Is that acceptable? The optimal amount of chip diversion is not zero. But I think this level of diversion would be a big deal, and the bigger concern is remote access.
I want to presume, for overdetermined reasons, that Sacks’s statement was written without due consideration or it does not reflect his actual views, and we would not actually make this level of dumb mistake where they could literally just swap the chips out for dummy chips. I presume we are planning to use vastly superior and more effective precautions against chip diversion and also have a plan for robust monitoring of compute use to prevent remote access diversion.
But how can we trust an administration to take such issues seriously, if their AI Czar is not taking this even a little bit seriously? This is not a one time incident. Similar statements keep coming. That’s why I spent a whole post responding to them.
David Sacks is also quoted extensively directly in the Bloomberg piece, and is repeatedly very dismissive of worried about diversion of chips or of compute, saying it is a fake argument and an easy problem to solve, and he talks about these as if they were reliable American allies in ways I do not believe are accurate.
Sacks also continues to appear to view winning AI to be largely about selling AI chips. As in, if G42, an Abu Dhabi-based AI firm, is using American AI chips, then it essentially ‘counts as American’ for purposes of ‘winning,’ or similar. I don’t think that is how this works, or that this is a good use of a million H100s. Bloomberg reports 80% of chips headed to the UAE would go to US companies, 20% to G42.
I very much want us to think about the actual physical consequences of various actions, not what those actions symbolize or look like. I do think, despite everything else, it is a very good sign that David Sacks is ‘urging people to read the fine print.’ This is moderated by the fact that we do not have the fine print, so we can’t read it. The true good news there requires one to read all that fine print, and one also should not assume that the fine print will get implemented. Nor do we yet have access to what the actual fine print says, so we cannot read it.
Semianalysis Defends the Deal
Dylan Patel and others at Semianalysis offer a robust defense of the deal, saying clearly that ‘America wins’ and that this benefits American AI infrastructure suppliers on all levels, including AI labs and cloud providers.
They focus on three benefits: money, tying KSA/UAE to our tech stack, and electrical power, and warn of the need for proper security, including model weight security, a point I appreciated them highlighting.
Those seem like the right places to focus, and the right questions to ask. How much of their money is really up for grabs and how much does it matter? To what extent does this meaningfully tie UAE/KSA to America and how much does that matter? How much do we need their ability to provide electrical power? How will the security arrangements work, will they be effective, and who will effectively be in charge and have what leverage?
Specifically, on their three central points:
They call this macro, but a better term would be money. UAE and KSA (Saudi Arabia) can make it rain, a ‘trillion-dollar floodgate.’ This raises two questions.
Question one: Was American AI ‘funding constrained’? The big tech companies were already putting in a combined hundreds of billions a year. Companies like xAI can easily raise funds to build giant data centers. If Google, Amazon, Apple, Meta or Microsoft wanted to invest more, are they really about to run out of available funding? Are there enough more chips available to be bought to run us out of cash?
Semianalysis seems to think we should be worried about willingness of American companies to invest here and thinks we will have trouble with the financing.
I am not convinced of this. Have you seen what these companies (don’t have to) pay on corporate bonds? Did we need to bring in outside investors? Should we even want to, given these investments look likely to pay off?
This is a major crux. If indeed American big tech companies are funding constrained in their AI investments, then the money matters a lot more. Whereas if we were already capable of buying up all the chips, that very much cuts the other way.
Question two: As we discussed earlier, is the trillion-dollar number real? We keep seeing these eye-popping headline investment numbers, but they don’t seem that anchored to reality, and seem to include all forms of investment including not AI, although of course other foreign direct investment is welcome.
Do their investments in US datacenters mean anything, and are they even something we want, given that the limiting factor driving all this is either constraints on chip availability or on electrical power? Will this be crowding out other providers?
If these deals are so positive for American tech companies, why didn’t the stock market moves reflect this? No, I will not accept ‘priced in.’
They call this geopolitical, that UAE and KSA are now tied to American technology stacks.
As they say, ‘if Washington enforces tight security protocols.’ We will see. David Sacks is explicitly dismissing the need for tight security protocols.
Classically, as Trump knows well, when the bank loans you a large enough amount and you don’t pay it back, it is the bank that has the problem. Who is being tied to whose stack? They will be able to at least cut the power any time. It is not clear from public info what other security will be present and what happens if they decide to turn on us, or use that threat as leverage. Can they take our chips and their talents elsewhere?
This can almost be looked at as a deal with one corporation. G42 seems like it’s going to effectively be on the UAE side of the deal, and it is going to have a lot of chips in a lot of places. A key question is, to what extent do we have the leverage on and control over G42, and to what extent does this mean they will act as a de facto American tech company and ally? How much can we trust that our interests will continue to align? Who will be dependent on who? Will our security protocols extend to their African and European outposts?
Why does buying a bunch of our chips tie them into the rest of our stack? My technical understand is that it doesn’t. They’re only tied to the extent that they agreed to be tied as part of the deal (again, details unknown), and they could swap out that part at any time. In my experience you can change which AI your program uses by changing a few lines of code, and people often do.
It is not obvious why KSA and UAE using our software or tech stack is important to us other than because they are about to have all these chips. These aren’t exactly huge markets. If the argument is they have oversized effect on lots of other markets, we need to hear this case made out loud.
Seminanalysis points out China doesn’t even have the capacity to sell its own AI chips yet. And I am confused about the perspectives here on ‘market share’ and the implied expectations about customer lock-in.
They call this infrastructure, I’d simply call it (electrical) power. This is the clearly valuable thing we are getting. It’s rather crazy that ‘put our most strategic asset except maybe nukes into the UAE and KSA’ was chosen over ‘overrule permitting rules and build some power plants or convince one of our closer allies to do it’ but here we are.
So the question here is, what are the alternatives? How acute is the shortage going to be and was there no one else capable of addressing it?
Also, even if we do have to make this deal now, this is screaming from the rooftops, we need to build up more electrical power everywhere else now, so we don’t have this constraint again in the future.
Semianalysis also raises the concern about model weight security, but essentially think this is solvable via funding work to develop countermeasures and use of red teaming, plus defense in depth. It’s great to see this concern raised explicitly, as it is another real worry. Yes, we could do work to mitigate it and impose good security protocols, and keep the models from running in places and ways that create this danger, but will we? I don’t know. Failure here would be catastrophic.
Understanding the China Hawks
There are also other concerns even if we successfully retain physical and digital control over the chips. The more we place AI chips and other strategic AI assets there, the more we are turning UAE, Saudi Arabia and potentially Qatar into major AI players, granting them leverage I believe they can and will use for various purposes.
David Sacks continues to claim to not understand that others think that ‘winning AI’ is mostly not about who gets to sell chips, who uses our models and picks up market share, or about superficially ‘winning’ ‘deals.’
He not only thinks it is about market penetration, he can’t imagine an alternative. He doesn’t understand that many, including myself, this is about who has compute and who gets superintelligence, and about the need for proper security.
David Sacks: I’m genuinely perplexed how any self-proclaimed “China Hawk” can claim that President Trump’s AI deals with UAE and Saudi Arabia aren’t hugely beneficial for the United States. As leading semiconductor analyst Dylan Patel observed, these deals “will noticeably shift the balance of power” in America’s favor. The only question you need to ask is: does China wish it had made these deals? Yes of course it does. But President Trump got there first and beat them to the punch.
Sam Altman: this was an extremely smart thing for you all to do and i’m sorry naive people are giving you grief.
Tripp Mickle and Ana Swanson (NYT): One Trump administration official, who declined to be named because he was not authorized to speak publicly, said that with the G42 deal, American policymakers were making a choice that could mean the most powerful A.I. training facility in 2029 would be in the United Arab Emirates, rather than the United States.
But Trump officials worried that if the United States continued to limit the Emirates’ access to American technology, the Persian Gulf nation would try Chinese alternatives.
The hawks are concerned, because the hawks largely do not think that the key question is who will get to sell chips, but rather who gets to buy them and use them. This is especially true given that both America and China are producing as many top AI chips as they can, us far more successfully, and there is more than enough demand for both of them. One must think on the margin.
Given that so many China hawks are indeed on record doubting this deal, if you are perplexed by this I suggest reading their explanations. Here is one example.
Tripp Mickle and Ana Swanson (NYT): Mr. Goodrich said the United States still had the best A.I. engineers, companies and chips and should look for ways to speed up permitting and improve its energy grid to hold on to that expertise. Setting up some of the world’s largest data centers in the Middle East risks turning the Gulf States, or even China, into A.I. rivals, he said.
“We’ve seen this movie before and we should not repeat it,” Mr. Goodrich said.
Sam Winter-Levy, a fellow at the Carnegie Endowment for International Peace, said the huge chip sales did “not feel consistent with an America First approach to A.I. policy or industrial policy.”
“Why would we want to offshore the infrastructure that will underpin the key industrial technology of the coming years?” he asked.
This does not seem like a difficult position to understand? There are of course also other reasons to oppose such deals.
Here is Jordan Schneider of China Talk’s response, in which he is having absolutely none of it, explicitly rejecting that either America or China has chips to spare for this. rejecting that UAE and KSA are actual allies, not expecting us to follow through with reasonable security precautions, and saying if we wanted to do this anyway we could have held out for a better deal with more control than this, I don’t know why you would be confused how someone could have this reaction based on the publicly available information:
Jordan Schneider: It’s going to cannibalize US build-out and leave the world with three independent power-centers of AI hardware where we could’ve stuck to our guns, done more power generation at home, and only had China to deal with not these wild-card countries that are not actual allies. If this really is as important as we believe, why are we letting these countries and companies we deeply distrust get access to it?
The Gulf’s BATNA wasn’t Huawei chips, it was no chips. Whatever we’re trying to negotiate for, we can play harder to get. BIS can just say they can’t buy Ascends and it’s not like there’s enough capacity domestically in China to service global demand absent the TSMC loophole they charged through. Plus, we’re offering to sell them 10× the chips that Huawei could conceivably sell them anytime soon even if they use the TSMC-fabbed wafers.
Where’s the art-of-the-deal energy here? Right now I only see AMD and NVDA shareholders as well as Sama benefiting from all of this. I thought we wanted to raise revenue from tariffs? Why not charge 3× the market rate and put the premium into the US Treasury, some “Make America Great Again” industrial-development fund, use it to triple BIS’ budget so they can actually enforce the security side, put them on the hook for Gaza…I don’t know literally anything you care about. How about a commitment not to invest in Chinese tech firms? Do we still care about advanced logic made in America? How about we only let them buy chips fabbed in the US, fixing the demand-side problem and forcing NVDA to teach Intel how to not suck.
Speaking of charging through loopholes, all of the security issues Dylan raises in his article I have, generously, 15 % confidence in USG being able to resolve/resist industry and politicians when they push back. If it’s so simple to just count the servers, why hasn’t BIS already done it / been able to fight upstream industry lobbying to update the chips-and-SME regs to stop Chinese build-outs and chip acquisition? What happens when the Trump gets a call from the King when some bureaucrat is trying to stop shipments because they see diversion if they ever catch it in the first place?
Why are we doing anything with G42 again? Fine, if you really decide you want to sell chips to the UAE, at the very least give American hyperscalers the off-switch. It’s not like they would’ve walked away from that offer! America has a ton to lose in the medium term from creating another cloud provider that can service at scale, saying nothing of one that has some deeply-discomforting China ties pretty obvious even to me sitting here having never gotten classified briefings on the topic.
Do the deal’s details and various private or unvoiced considerations make this deal better than it looks and answer many of these concerns? Could this be sufficient that, if looked at purely through the lens of American strategic interests, this deal was a win versus the salient alternatives? Again: That is all certainly possible!
Our negotiating position could have been worse than Jordan believes. We could have gotten important things for America we aren’t mentioning yet. The administration could have limited room to maneuver including by being divided against itself or against Congress on this. On the flip side, there are some potentially uncharitable explanations for all of this, that would be reasonable to consider.
Rhetoric Unbecoming
Instead of understanding and engaging with such concerns and working to allay them, Sacks has repeatedly decided to make this a mask off moment, and engage in a response that I would expect on something like the All-In Podcast or in a Twitter beef, but which is unbecoming of his office and responsibilities, with multiple baseless vibe and ad hominem attacks at once that reflect that he either is willfully ignorant of the views, goals and beliefs of those he is attacking and even who they actually are, or he is lying and does not care, or both, and a failure to take seriously the concerns and objections being raised. Here is another illustration of this:
David Sacks (May 17): After the Sam Bankrun-Fraud fiasco, it was necessary for the Effective Altruists to rebrand. So they are trying to position themselves as “China Hawks.” But their tech deceleration agenda is the same, and it would cost America the AI race with China.
There are multiple other people I often disagree with on important questions but whom I greatly respect who are working on in administration on AI policy. There are good arguments you can make in defense of this deal. Instead of making those arguments in public, we repeatedly get this.
This is what I call Simulacra Level 4. Everything Sacks says seems to be about vibes and implications first and actual factual claims a distant second at best. He doesn’t logically say ‘all so-called China hawks who don’t agree with me are secret effective altruists in trench coats and also decels who hate all technology and all of humanity and also America,’ but you better believe that’s the impression he’s going for here.
Could China Have ‘Done This Deal’?
Would China have preferred to ‘do this deal’ instead? That at best assumes facts, and arguments, not in evidence. It depends what they would get out of such a deal, and what we’re getting out of ours, and also the security arrangements and whether we’ve formed a long lasting relationship in which we hold the cards.
I’m also not even sure what it would mean for China to have ‘done this deal,’ it does not have what we are offering. Semianalysis says they don’t have similar quantities of chips to sell, and might not have any, nor are their chips of similar quality.
I do agree China would have liked to ‘do a deal’ in some general sense, where they bring UAE/KSA into their orbit, on AI and otherwise, although they don’t need access to electrical power. More capital and friends are always helpful. It’s not clear what that deal would have looked like.
One must again emphasize: There is a lot that we do not know, that matters a lot, or even that has yet to be worked out. Diplomacy often must be done in private. It is entirely possible that there is more information, or there are more arguments and considerations, behind the scenes that justifies what is being done, and that the final deal here is a win and makes sense.
Tyler Cowen: Of course Saudi and the UAE have plenty of energy, including oil, solar, and the ability to put up nuclear quickly. We can all agree that it might be better to put these data centers on US territory, but of course the NIMBYs will not let us build at the required speeds. Not doing these deals could mean ceding superintelligence capabilities to China first. Or letting other parties move in and take advantage of the abilities of the Gulf states to build out energy supplies quickly.
Energy and ability to overcome NIMBYs is only that which is scarce because America is refusing to rise to this challenge and actually enable more power generation. Seriously, is there nowhere in America we can make this happen at scale? If we wanted to, we could do this ourselves easily. We have the natural gas, even if nuclear would be too slow to come online. It is a policy choice not to clear the way. And no, I see zero evidence that we are pulling out the stops here and coming up short.
I think this frame is exactly correct – that this deal makes sense if and only if all of:
The security deal is robust and we retain functional control over where the compute goes.
We trust our friends here to remain our friends at a reasonable price.
We counterfactually would not have been able to buy these chips and build data centers to power these chips.
As far as I can tell China already has all the power it needs to power any AI chips it can produce, it is using them all, and its chip efforts are not funding constrained.
So for want of electrical power, and for a few dollars, we are handing over a large amount of influence over the future to authoritarian powers with very different priorities and values?
Tyler Cowen: In any case, imagine that soon the world’s smartest and wisest philosopher will soon again be in Arabic lands.
We seem to be moving to a world where there will be four major AI powers — adding Saudi and UAE — rather than just two, namely the US and China. But if energy is what is scarce here, perhaps we were headed for additional AI powers anyway, and best for the US to be in on the deal?
Who really will have de facto final rights of control in these deals? Plug pulling abilities? What will the actual balance of power and influence look like? Exactly what role will the US private sector play? Will Saudi and the UAE then have to procure nuclear weapons to guard the highly valuable data centers? Will Saudi and the UAE simply become the most powerful and influential nations in the Middle East and perhaps somewhat beyond?
Yes. Those are indeed many of the right questions, once you think security is solid. Who is in charge of these data centers in the ways that matter? Won’t they at minimum have the ability to cut the power at any time? Who gets to decide where the compute goes? What are they going to do with all this leverage we are handing them?
Is this what it means to have the future be based on American or Democratic values? Do you like ‘the values’ of the UAE and Saudi Arabian authorities?
Tyler Cowen: I don’t have the answers to those questions. If I were president I suppose I would be doing these deals, but it is very difficult to analyze all of the relevant factors. The variance of outcomes is large, and I have very little confidence in anyone’s judgments here, my own included.
Few people are shrieking about this, either positively or negatively, but it could be the series of decisions that settles our final opinion of the second Trump presidency.
The administration thinks that the compute in question will remain under the indefinitely control of American tech companies, to be directed as we wish.
Sriram Krishnan: Reflecting on what has been an amazing week and a key step in global American AI dominance under President Trump.
These Middle East AI partnerships are historic and this “AI diplomacy” will help lock in the American tech stack in the region, help American companies expand there while also building infrastructure back in the U.S to continue expanding our compute capacity.
This happens on top of rigorous security guarantees to stop diversion or unauthorized access of our technology.
More broadly this helps pull the region closer to the U.S and aligns our technological interests in a very key moment for AI.
It’s a very exciting moment and a key milestone.
I hope that they are right about this, but I notice that I share Tyler’s worry that they are wrong.
The justification given for rescinding the Biden diffusion rules is primarily that failure to do this would have ‘weakened diplomatic relations with dozens of countries by downgrading them to second-tier status.’
But, well, not to reiterate everything I said last week, but on that note I have news.
One, we’re weakening diplomatic relations with essentially all countries in a series of unforced errors elsewhere, and we could stop.
Two, most of the listed tier two countries have always had second-tier status. There’s a reason Saudi Arabia isn’t in Five Eyes or NATO. We can talk price about which countries should have which status, but no our relations are not all created equal, not when it comes to strategically vital national interests and to deep trust. I don’t share Sacks’s stated view that these are some of our closest and most trustworthy allies. Why does this administration seem to always want to make its deals mostly with authoritarian regimes, usually in places where Trump has financial ties?
Tripp Mickle and Ana Swanson (NY Times): The announcements of the two deals follow reports that $2 billion has flowed to Trump companies over the last month from the Middle East, including a Saudi-backed investment in Trump’s cryptocurrency and plans for a new presidential airplane from Qatar.
There’s always Trust But Verify. The best solution, if you can’t trust, is often to set up things so that you don’t have to. This can largely be done. Will we do it? And what will we get in return? What is announced mostly seems to be investments and purchases, that what we are getting are dollars, and Bloomberg is skeptical of the stated dollar amounts.
At Best A Second Best Solution
This deal is very much not a first best solution. It is, at best, a move that we are forced into on the margin due to our massive unforced errors in a variety of other realms. Even if it makes sense to do this, it makes even more sense to be addressing and fixing those other critical mistakes.
Electrical power is the most glaring in the context of this particular. There needs to be national emergency level focus on America’s inability to build electrical power capacity. Where are the special compute zones? Where are the categorical exemptions? Where is DOGE with regard to the NRC? Where is the push for real reform on any of these fronts? Instead, we see story after story of Congress actively moving to withdraw even the supports that are already there, including plans to outright abrogate contracts on existing projects.
The other very glaring issue is trade policy. If we think it is this vital to maintain trade alliances and open up markets, and maintaining market share, why are we otherwise going in the opposite direction? Why are we alienating most of our allies? And so on.
The argument for this deal is, essentially, that it must be considered in isolation. That other stuff is someone else’s department, and we can only work with what we have. But this is a very bitter pill to be asked to swallow, especially as Sacks himself has spoken out quite loudly in favor of many of those same anti-helpful policies, and the others he seems to be sitting out. You can argue that he needs to maintain his political position, but if that also rules out advocating for electrical power generation and permitting reform, what are we even doing?
If we swallow the entire pill, and consider these deals only on the margin, without any ability to impact any of our other decisions, and only with respect to ‘beating China’ and ability to ‘win the AI race,’ and assume fully good faith and set aside all the poor arguments and consider only the steelman case, we can ask: Do these deals help us?
I believe that such a deal is justifiable, again on the margin and regarding our position with respect to China, if and only if ALL of the following are true:
Security arrangements are robust, the chips actually do remain under our physical control and we actually do determine what happens with the compute. And things are set up such that America retains the leverage, and we can count on UAE/KSA to remain our friends going forward.
This was essentially the best deal we could have gotten.
This represents a major shift in our or China’s ability to stand up advanced AI chips, because for the bulk of these chips either Big Tech would have run out of money, or we would have been unable to source the necessary electrical power, or China has surplus advanced AI chips I was not previously aware of and no way to deploy them.
Entering into these partnerships is more diplomatically impactful, and these friendships are more valuable, than they appear to me based on public info.
