Joe Walker has another typically great podcast out, this time chatting with Andrew Charlton on how Australia should handle the rise of the artificial intelligence boom. The two-hour podcast obviously touches on lots of different things. But my big takeway was the question of how Australia would benefit from AI models.
The paradoxical aspect of AI data centres is that local benefits can actually be quite minimal. While they do use some local labour, particularly around construction, they are mostly comprised of imported high end chops. So once they’re up and humming the benefits can be quite minor. They don’t hire many people, so they pay very little payroll tax. In the case of any of the main frontier labs, they would almost certainly need to license the weights from their U.S. owners, which would limit the profits of any domestic firm, and so they would generate pretty minimal profits locally, and thus pay little company tax. They would likely end up exporting a lot of compute around the world, which would limit the amount of GST they would pay (which only applies to local consumption).
And while they will be buying a boatload of Australian power, much of the additional supply is likely to come in the form of imported capital goods (liek the chips themeslves), such as solar panels and batteries. You could imagine the benefits there will also be similarly modest. And the data centres themselves comprise almost entirely imported AI chips. So while the AI data centre boom may boost our GDP statistics, I suspect the direct benefits of their role here could be quite minimal.
Tax and grab
Other places around the world have grappled with this problem by taxing the data centres themselves. Loudoun County in Virginia, one of the epicentres of the AI data centre boom, has a 3% tax on all business equipment. So they tax all the GPUs installed at around 3% a year, which to my untrained ear sounds like a criminally large and inefficient tax on capital goods. But it’s clearly not enough to scare off a tsunami of AI data centres, which has helped make it one of the wealthiest localities in America.
So this is one obvious approach Australia could take to make sure locals benefit; a new tax to raise revenue. And I think if it was negotiated now, while the data centre boom is just beginning, it would be fairly straightforward to implement.
However, according to this interview, that’s not what the government is doing. They seem to be taking more of a path whereby data centres located in Australia will be asked to provide benefits in other ways, as opposed to just handing over slabs of cash to the taxpayer.
Compute for Access
One idea raised by Joe, although not necessarily signed on to by Andrew, was this idea of model access in exchange for data centre approvals. The idea being that Australia would get better access to AI models if we managed to build more data centres locally.
I’m quite sceptical this will work. I think at the end of the day, the determining factor in whether access to frontier models is granted is almost entirely going to be driven by U.S. government policy and whether you’re trusted with them. Australia’s on relatively good ground here, because we are a Five Eyes partner and highly trusted. But that wasn’t enough to stop us being locked out of Fable by the Trump administration. And I don’t think us having hosted a dozen data centres for Anthropic would have changed that. At the end of the day, the law’s the law, and no amount of co-investment scheme is going to change that.
Perhaps you can imagine things being a bit different if frontier models were only open to trusted partners, as long as they BYOd their own compute. But I think this is an unlikely scenario. At the end of the day, a tightly restricted model is never going to be a big draw on computational resources, because even if it’s a large model that on its own takes up a lot of GPUs, if it’s restricted to just the U.S. government and its close allies, the number of users will always be very small compared to commercially available models.
So I don’t think compute-for-model-access is going to work out.
Some other options they suggested were offering subsidised compute rates for Australian firms and universities, which to me sounds needlessly complicated. How do we know which models Australian firms or universities might want to have access to in the future? And how do we even define what an Australian firm is? Would BYD Australia be able to get access to Anthropic models under this scheme in order to better sell electric cars in Australia? How would you design a scheme that prevented that from happening? It seems very tricky.
Another idea is hosting our own data centres so we can run or train our own localised versions of AI models. To be honest, I’m also pretty sceptical this is going to be needed. While I can see the case for a culture that is very different from the American one wanting to post-train an open-source model in order to fit some particular local task, the reality is Australia and America are incredibly similar culturally, and I highly doubt it would ever make sense for us to want to train our own models.
Even leaving aside copyright concerns, just how much need is there for an Australian-centric model? Bluey has become one of the most streamed shows on American TVs. I’m sure ChatGPT can provide a decent recipe for damper. So I think the scope for locally trained or post-trained models while remain marginal.
Liquidity is the ultimate hedge
And even putting aside all those concerns, I just think the uncertainty in this area is incredibly large. We don’t know whether AI will turn out to be a bust — unlikely but possible — or which firms will end up producing the best models in the next year or five years, which is highly uncertain, or what they’ll be used to do. That makes it nigh on impossible for any government to negotiate the right terms for some sort of access-for-data-centre deal.
In my mind, the obviously superior answer is to replicate something like what Loudoun County does. Tax GPUs when they’re installed at some modest but sustainable rate, approve as many data centres as we physically can, watch the money pour in, and use that money to buy whatever AI services — or indeed even normal goods and services — we think we need to better weather the future.
It may be that we spend the money just to buy the entire country a subscription to Claude. Or maybe we spend it on retraining workers to do non-AI-substitutable tasks and jobs. The benefit of taking the cash upfront instead of nebulous promises of access to AI is that money will almost always be able to buy access in the future, and we can change what we spend that money on as circumstances change, which is almost surely going to happen.
ntres is that local benefits can actually be quite minimal. While they do use some local labour, particularly around construction, they are mostly comprised of imported high end chops. So once they’re up and humming the benefits can be quite minor. They don’t hire many people, so they pay very little payroll tax. In the case of any of the main frontier labs, they would almost certainly need to license the weights from their U.S. owners, which would limit the profits of any domestic firm, and so they would generate pretty minimal profits locally, and thus pay little company tax. They would likely end up exporting a lot of compute around the world, which would limit the amount of GST they would pay (which only applies to local consumption).
And while they will be buying a boatload of Australian power, much of the additional supply is likely to come in the form of imported capital goods (liek the chips themeslves), such as solar panels and batteries. You could imagine the benefits there will also be similarly modest. And the data centres themselves comprise almost entirely imported AI chips. So while the AI data centre boom may boost our GDP statistics, I suspect the direct benefits of their role here could be quite minimal.
Tax and grab
Other places around the world have grappled with this problem by taxing the data centres themselves. Loudoun County in Virginia, one of the epicentres of the AI data centre boom, has a 3% tax on all business equipment. So they tax all the GPUs installed at around 3% a year, which to my untrained ear sounds like a criminally large and inefficient tax on capital goods. But it’s clearly not enough to scare off a tsunami of AI data centres, which has helped make it one of the wealthiest localities in America.
So this is one obvious approach Australia could take to make sure locals benefit; a new tax to raise revenue. And I think if it was negotiated now, while the data centre boom is just beginning, it would be fairly straightforward to implement.
However, according to this interview, that’s not what the government is doing. They seem to be taking more of a path whereby data centres located in Australia will be asked to provide benefits in other ways, as opposed to just handing over slabs of cash to the taxpayer.
Compute for Access
One idea raised by Joe, although not necessarily signed on to by Andrew, was this idea of model access in exchange for data centre approvals. The idea being that Australia would get better access to AI models if we managed to build more data centres locally.
I’m quite sceptical this will work. I think at the end of the day, the determining factor in whether access to frontier models is granted is almost entirely going to be driven by U.S. government policy and whether you’re trusted with them. Australia’s on relatively good ground here, because we are a Five Eyes partner and highly trusted. But that wasn’t enough to stop us being locked out of Fable by the Trump administration. And I don’t think us having hosted a dozen data centres for Anthropic would have changed that. At the end of the day, the law’s the law, and no amount of co-investment scheme is going to change that.
Perhaps you can imagine things being a bit different if frontier models were only open to trusted partners, as long as they BYOd their own compute. But I think this is an unlikely scenario. At the end of the day, a tightly restricted model is never going to be a big draw on computational resources, because even if it’s a large model that on its own takes up a lot of GPUs, if it’s restricted to just the U.S. government and its close allies, the number of users will always be very small compared to commercially available models.
So I don’t think compute-for-model-access is going to work out.
Some other options they suggested were offering subsidised compute rates for Australian firms and universities, which to me sounds needlessly complicated. How do we know which models Australian firms or universities might want to have access to in the future? And how do we even define what an Australian firm is? Would BYD Australia be able to get access to Anthropic models under this scheme in order to better sell electric cars in Australia? How would you design a scheme that prevented that from happening? It seems very tricky.
Another idea is hosting our own data centres so we can run or train our own localised versions of AI models. To be honest, I’m also pretty sceptical this is going to be needed. While I can see the case for a culture that is very different from the American one wanting to post-train an open-source model in order to fit some particular local task, the reality is Australia and America are incredibly similar culturally, and I highly doubt it would ever make sense for us to want to train our own models.
Even leaving aside copyright concerns, just how much need is there for an Australian-centric model? Bluey has become one of the most streamed shows on American TVs. I’m sure ChatGPT can provide a decent recipe for damper. So I think the scope for locally trained or post-trained models while remain marginal.
Liquidity is the ultimate hedge
And even putting aside all those concerns, I just think the uncertainty in this area is incredibly large. We don’t know whether AI will turn out to be a bust — unlikely but possible — or which firms will end up producing the best models in the next year or five years, which is highly uncertain, or what they’ll be used to do. That makes it nigh on impossible for any government to negotiate the right terms for some sort of access-for-data-centre deal.
In my mind, the obviously superior answer is to replicate something like what Loudoun County does. Tax GPUs when they’re installed at some modest but sustainable rate, approve as many data centres as we physically can, watch the money pour in, and use that money to buy whatever AI services — or indeed even normal goods and services — we think we need to better weather the future.
It may be that we spend the money just to buy the entire country a subscription to Claude. Or maybe we spend it on retraining workers to do non-AI-substitutable tasks and jobs. The benefit of taking the cash upfront instead of nebulous promises of access to AI is that money will almost always be able to buy access in the future, and we can change what we spend that money on as circumstances change, which is almost surely going to happen.