Ore Grade, AI and the $4.5 Billion Question at Bagdad

•

Topi Group

6-MINUTE READ · Oct 11, 2026

At a glance

  • Freeport-McMoRan is weighing roughly $4.5 billion, about 30% above its 2023 estimate, to more than double the concentrator at its Bagdad copper mine, with a decision due by year-end.
  • AI cannot change the grade in the ground, but it can change how much copper you recover, what counts as ore, and whether the next billion dollars performs.
  • At Bagdad, an AI model lifted mill throughput 10% in 2018 and avoided most of a $200 million expansion Freeport had been weighing, because operators adopted it and Freeport’s own team took it over.
  • Going fully autonomous forced the site to mature: standard work enforced by software, a data layer for the next wave of AI, new roles for more than 200 drivers, and a higher reliability bar.
  • The lesson for leaders is sequencing: sweat the existing assets, let automation reshape the operating model, and only then spend the capital.

In July, Freeport-McMoRan told investors that new concentrating facilities at its Bagdad mine in Arizona would cost about 30% more than the $3.5 billion it estimated in 2023; roughly $4.5 billion. A decision is due by year-end 2026. If it goes ahead, the project would add 200 to 250 million pounds (Mlbs) of copper a year and make Bagdad the second-largest copper mine in the United States.

Bagdad’s first mining claims were staked in 1882, and its first mill started up in 1928. Output peaked at 218 Mlbs in 2019 and was 149 Mlbs in 2025. So here is a fair question for anyone who believes, as we do, in AI-enabled operations: if the technology is so good, why does a mine that has used it for years still need to pour concrete?

Our answer: AI cannot change the grade in the ground. It can change how much of it you recover, what counts as ore, and whether the next billion dollars performs. Bagdad shows all three.

Grade is geology

Let’s be clear about the limit first. No model reverses geology. Lower grades mean more tonnes moved and milled for every pound of copper, and at Bagdad grade was not the only headwind. Freeport reported that understaffing at its U.S. operations held back metal output in 2022, and Bagdad was likely affected too. In addition, the concentrator’s capacity, now cited at 77,100 tonnes per day (tpd), is about 5% below the post-2018 peak of 81,600 tpd Freeport reported in 2019. Anyone who tells you an algorithm offset all of that is selling something.

But “AI cannot fix grade” is where the conversation should start, not where it ends. The interesting question is what a mature, lower-grade operation does before it reaches for capital.

First, sweat the asset

According to McKinsey, which worked with Freeport on what followed, the company was weighing a $200 million expansion at Bagdad in 2018 to lift production by 20%. Then copper fell from about $3.30 per pound in early June to about $2.75 per pound a month later, and the economics thinned. Instead, working with McKinsey, Freeport’s metallurgists built a model called TROI that sorted mill feed into seven ore types, rather than treating every tonne the same, and recommended plant settings for each.

In the fourth quarter of 2018, mill throughput rose 10% above the prior quarter, and recovery rose a full percentage point. By McKinsey’s account, the gain was worth about 20 million pounds of copper a year, and Freeport avoided most of the $200 million. Freeport’s own filings are consistent with that: the concentrator’s stated capacity rose from 75,000 tpd in its 2018 annual report to 81,600 tpd a year later. By November 2019, Freeport’s senior management was telling the Financial Times that the company had “set an aspirational goal of adding 200m pounds (90,000 tonnes) of copper from these initiatives with very little capital investment.”

The part that interests us most as operators is not the model. It is that metallurgists were accepting more than 80% of its recommendations within weeks, and that Freeport’s own metallurgists and data scientists went on to run it without the consultants. That is adoption, and adoption is the whole game. BHP’s Shingo Prize for its cathodes operation at Escondida, earned on an operating system of 14 practices, makes the same point from the other direction: the management system comes first, and the analytics ride on it.

The playbook did not stop at the concentrator. Across its Americas operations, Freeport has since turned data analytics and new technologies on leaching, recovering copper from material once treated as waste. That program reached a run rate of about 240 Mlbs of copper a year in late 2025, with 300 Mlbs targeted by the end of 2026. It is the closest thing to “fixing grade” we know of: not changing the rock, but changing what counts as ore.

Then, let autonomy force the site to mature

In 2025 Bagdad became the first major U.S. mine to run a fully autonomous haul fleet: 33 trucks designed to carry more than 95% of its haulage. Freeport has been candid about why. The site had trouble finding and retaining labor, and, in the company’s words, “We are implementing AHS at a site that is positioned to grow. Increasing the workforce to support manned haulage is not a viable option.”

Take that at face value. But we think the bigger payoff is what driverless haulage forces a mine to become. A manned mine absorbs variability through people’s judgment. An autonomous one cannot, so it has to make its discipline explicit. Bagdad ran the change in four phases, from an isolated test zone to full autonomy. It added a new command center, and two tunnels so light vehicles could reach work areas without entering the autonomous zone. Everyone who drives in the pit must complete training, including on a simulator built from real mine footage. And 217 haul truck drivers moved into other roles under a “Zero Job Loss” principle, about 80 of them onto shovels.

Four things happen when a site goes autonomous:

  1. Standard work gets enforced by software. Operational excellence programs spend years trying to get there with people alone.
  2. The data layer gets built. Every truck becomes a sensor stream, the foundation the next wave of AI needs.
  3. The constraint moves. Once haulage is predictable, the bottleneck shifts to loading, drilling or the mill, and the organization has to follow it.
  4. The reliability bar rises for everyone. Our team once stood in the dispatch room at a gold mine in South Carolina, and watched the fleet-management system drop out every time hard rain took down the site Wi-Fi. The trucks were manned, so nothing stopped. An autonomous fleet would have. We saw the same at a nickel operation in New Caledonia: two sites on opposite sides of the globe, running different fleet-management systems, with the same result. Autonomy makes weak infrastructure and loose process expensive, immediately.

Bagdad is not alone. Rio Tinto’s autonomous trucks in the Pilbara each ran an estimated 700 more hours a year than manned trucks, at about 15% lower load-and-haul unit cost, and the company held 2,000 career conversations in 2019. It has also said people, not technology, must be the focus of the change. Glencore has taken automated loaders and trucks underground at George Fisher in Queensland. And this year the Shingo Prize, one of the most rigorous awards for operational excellence culture, went to Vale’s Brucutu mine in Brazil for its autonomous haulage operation, which put its first driverless trucks to work in 2018 and now runs a fully autonomous fleet. Autonomy and an award-winning OpEx culture grew up together there.

The caveat: the forcing function works only if leaders treat autonomy as a change to the operating model. Bolted onto manned-era habits, it underdelivers.

Only then, spend the capital

Which brings us back to the $4.5 billion. It is tempting to argue that, without its technology, Freeport would have had to build years ago. The evidence does not support that, and we would rather say so than overclaim. TROI added about 20 Mlbs of copper a year; the expansion would add 200 to 250 Mlbs. The project needs a copper price of about $4.00 per pound, and Freeport realized less than $3.00 per pound in each year from 2017 to 2019. In 2022 the company expected to double Bagdad “in the 2026 timeframe.” Since then the timeline has slipped and the cost estimate has risen about 30%; nothing Freeport has disclosed suggests algorithms made the project unnecessary.

What technology did was subtler, and we think more valuable. It let Freeport squeeze its existing base first, for a fraction of the capital. It is weighing the big build now that copper sits well above that hurdle: the LME averaged $4.51 per pound in 2025. And the new plant, if approved, will land on a site that has already rebuilt its haulage, its data and its roles for scale. Newmont made a related bet at Boddington. When it approved a $150 million net investment in autonomous haulage in 2020, it projected an internal rate of return above 35% and at least two more years of mine life. It then converted 36 trucks in about seven months, on time and on budget, creating what it calls the gold industry’s first fully autonomous haul fleet.

That is what capital discipline looks like in practice: sequencing. Technology does not replace the capital. It raises the odds that the capital performs.

Four questions before the next capital request

If you run or govern a mature operation, these are worth asking before the next expansion reaches the board:

  1. What did we get out of our existing assets first, and can we show it in unit costs and recovery rather than pilots?
  2. Is automation changing our operating model, or is it bolted onto the old one?
  3. Who owns the models and the learning: our people, or a vendor or consultant?
  4. Before the new plant arrives, is the organization already running at the standard autonomy demands?

Grade sets the problem. The operating model decides the answer.

Sources

  1. Freeport-McMoRan company filings: 2018-2026; Bagdad autonomy updates, Jan 2025 and Oct 2025
  2. Newmont company filings: Boddington autonomous haulage release (Feb 2020); Boddington update (Oct 2021)
  3. E&MJ: Freeport plans to convert the Bagdad mine to autonomous haulage (Jan 2024); Boddington commissions 36 autonomous trucks in 7 months
  4. McKinsey, Inside a mining company’s AI transformation (Feb 2020)
  5. Investing News Network, citing the Financial Times (Nov 2019)
  6. Mining Weekly, US worker shortage denting Freeport-McMoRan’s copper output (Jan 2023)
  7. Global Mining Guidelines Group, Rio Tinto’s experience with automation and people (2021)
  8. Shingo Institute, Escondida | BHP (Mar 2023)
  9. Vale, Brucutu wins the Shingo Prize (Mar 2026)
  10. Glencore, George Fisher underground automation (Sep 2025)

Disclaimer: This article is based on Topi Group’s research using the publicly available sources cited above, current as of the publication date. The views and opinions expressed are our own and do not represent those of any company mentioned. Topi Group is not affiliated with, and has not been compensated by, any company mentioned. Nothing in this article constitutes investment, financial or legal advice. We make no representation as to the completeness or accuracy of third-party information, and we undertake no obligation to update or revise this article after publication, whether as a result of new information, future events or otherwise.