Special — Ramp-up Curves
Both hosts are synthetic voices and the script is AI-generated. Every paper cited was verified against a primary source (DOI or arXiv ID) before publication, but the commentary is not that of practising researchers.
Traci: Christmas Day, 1998. Western Australia, about an hour's drive northwest of Kalgoorlie. A brand-new nickel plant called Cawse produces its very first product. A bag of cobalt sulphide.
Joao: A bag. Singular.
Traci: One bag. Nickel cathode follows in January. It's one of three laterite plants starting up in Western Australia at almost the same moment, all built around pressure acid leaching, all financed by high-yield bonds, and all very publicly watched. Cawse gets to its design production level in May 2000, about twenty months after commissioning started.
Joao: Twenty months doesn't sound like a disaster for something that new.
Traci: It isn't. Among the three, it's the one that went best. And in March 2001, ten months after reaching design, the company that built it was placed into receivership by its bondholders. The plant was sold to another company.
Joao: So the plant worked and the owner went broke anyway.
Traci: The plant eventually worked, and the owner couldn't survive the time it took. That gap is what this episode is about. How fast do new mines and plants actually reach nameplate? How does that compare with what the feasibility study promised? And when the two disagree, who pays?
Joao: This is our second special, and it follows straight on from the last one, the mine-plan benchmarking episode.
Traci: It does. Last time the frame was Bent Flyvbjerg and Dan Gardner's book, How Big Things Get Done: the idea that you shouldn't build a forecast purely from the inside, from your own project's details, but anchor it on what a reference class of comparable projects actually achieved, the outside view, and only then adjust. Ramp-up is where that idea bites hardest, because a feasibility study has a ramp-up schedule in it, a month-by-month or year-by-year line, and there's an industry with decades of records of what really happened. Inside view, meet outside view.
Joao: And I'll tell you now where I'm going to push back. I'm not convinced the outside view here is as clean as people make it sound. I think some of what passes for a law of nature is really a budgeting habit.
Traci: Good. Hold that. You'll get your turn.
Joao: Before we go, the honest disclosure. What have we actually read?
Traci: About sixty documents. The original McNulty papers from 1998, 2004 and 2014. Two RAND Corporation reports from the eighties on process plant startups. A pile of start-up case histories written by the people who ran them: Kumtor, Cawse, Century, Cerro Verde, Constancia, Carmen de Andacollo, Mount Milligan and others. The feasibility-accuracy literature. And four technical reports to test against. What we haven't read is important too. Terry McNulty published updated curves in 2024 and 2025 in Springer conference volumes, and we don't have those chapters. What we do have is a blog series from the engineering firm Hatch, called The Ramp-up Series, and the first and fifth parts are written by Britt MacKinnon, Terry McNulty and Phillip Mackey. Whenever we talk about the modern update, that's our source: those Hatch posts, not the Springer papers.
Joao: Okay. So start me at the beginning. Who is McNulty and what's the curve?
Traci: Terry McNulty is a metallurgical engineer. He spent about twenty years at Anaconda, later ran a consultancy in Tucson, and in October 1998 he published a short paper in Mining Engineering called Developing Innovative Technology. In it he takes forty-one processing plants built between about 1965 and 1996. Flotation concentrators, smelters, chemical plants, leach plants. For each one he asks a single question: what percent of design capacity was it running at six months after start-up, twelve months, eighteen, thirty-six?
Joao: Just that one number over time.
Traci: Just that. And he borrowed the method from RAND, which we'll come to. He plots them all, sees that they fall into bands, and averages each band into a curve. Four curves. Series 1 is the good one: it's essentially at design capacity by the end of the first year. Series 4 is the bad one: it's still only a little over half of design three years in.
Joao: How many plants landed in each?
Traci: Twenty-three of the forty-one were Series 1. So a bit more than half got there in about a year. The other eighteen were spread across Series 2, 3 and 4, and the Series 4 group is the one people remember, because of the seven plants in it, three closed within thirty-six months of starting up.
Joao: So nearly half of the Series 4 plants didn't just run badly, they shut.
Traci: They shut. And the part of the paper that made it last isn't really the curves, it's the description of what each group had in common. Series 1 plants used mature technology, equipment similar in size and duty to things that had worked before, and properly piloted anything risky. As you move down, you start seeing first licensees of a new process, prototype equipment, pilot work done on unrepresentative samples, ore mineralogy not fully understood. And by Series 4 you get flowsheets with prototype equipment in two or more critical steps, process chemistry that was misunderstood, and one line I think every reviewer should have on a sticky note: equipment downsized, or design criteria made less conservative, in response to projected cost overruns.
Joao: The value-engineering trap.
Traci: The value-engineering trap. And McNulty's own conclusion, which he repeats in 2004, is a bit surprising. He went in expecting innovation itself to be the villain. What he found was that the care with which the technology was developed and implemented mattered more than how new it was.
Joao: Fine. Now you said he borrowed the method from RAND. RAND as in the Santa Monica think tank?
Traci: The same one. In the late seventies the US Department of Energy was worried about synthetic fuels plants, oil shale and coal liquefaction, very novel, very expensive, and asked RAND to work out why first-of-a-kind process plants kept blowing their budgets and underperforming. RAND built a confidential database from oil, chemical and minerals companies. Kennecott Copper is one of the firms that agreed to be named. And two of their reports are, I'd argue, the real reference-class backbone for this whole topic.
Joao: Why those and not McNulty?
Traci: Because they did something McNulty didn't. They compared plan against actual. The 1981 report, by Ed Merrow, Kenneth Phillips and Christopher Myers, looked at forty-four pioneer plants. It points out that financial models usually assumed a plant would run at eighty-five to ninety-five percent of nameplate. Then it measured what the plants actually did in their second six months, months seven to twelve. Twenty-two of forty-three missed eighty-five percent.
Joao: Half.
Traci: Half. But here's the number I'd want everybody in mining to remember from that report. They split out plants that process solids: solid feed or solid product. There were fifteen of them. In months seven to twelve those fifteen averaged forty-nine percent of design. Only two of the fifteen reached eighty-five. The liquid and gas plants averaged eighty-four.
Joao: And a mine is a solids plant. That's literally the job.
Traci: That's the job. Now, these were pioneer plants, so it's the worst end of the distribution. But the gap between solids and fluids is the thing. RAND's explanation was that for the solids plants, only about a quarter of the heat and mass balance equations could be based on data from earlier commercial plants. For the fluid plants it was about seventy percent. You know much less about how rock behaves at scale than you think you do.
Joao: And the second RAND report?
Traci: That's the one I promised. 1986, Myers, Shangraw, Devey and Hayashi, Understanding Process Plant Schedule Slippage and Startup Costs. Fifty-six plants, and this is where the comparison with plans gets really pointed. They plotted planned start-up time against actual start-up time for each plant. If plans were any good, the dots would sit on a forty-five-degree line.
Joao: And they didn't.
Traci: They were all over the place. The report's own wording is that knowing the planned startup time tells us very little about the time actually required. Planned start-ups clustered around a few months. The actual start-up time had a median of four months and a mean of eight, and that's with long ones capped at thirty months so they didn't swamp the statistics. And remember the definition here: start-up ends when the plant reaches fairly steady operation or is handed to the operating crew, not when it hits design capacity. So it flatters the plants.
Joao: Median four, mean eight. So a long tail.
Traci: A long tail. And then the line that, for me, is the single most useful sentence in the whole reference class. Among the pioneer plants fed with solids, none of the eleven met their startup schedules. Not one.
Joao: Wait, though. I've seen that RAND study quoted, and not like that. Didn't Hatch use it in their blog series?
Traci: They did, in Part 2, and I'm glad you raised it, because it's a nice example of how numbers drift. The Hatch post says that when ramp-up time increased from four months to eight months, ramp-up costs rose from three point six percent to five point five percent of construction capital.
Joao: That sounds like a clean before-and-after.
Traci: It sounds like one, but it isn't. Four and eight are the median and the mean of start-up time across the database. Three point six and five point five are the median and the mean of start-up cost as a share of capital, across a slightly different set of plants. They're two summary statistics of two distributions, laid side by side. RAND never says that doubling the start-up time added two percent to costs.
Joao: So the outside view gets an inside-view treatment on the way through.
Traci: It gets a bit of polish, yes. The underlying RAND finding is stronger anyway. First-of-a-kind plants in their data spent around nine percent of capital on start-up, against one or two percent for the others. And their regression gives you something you can actually use: roughly ten to eleven extra weeks of start-up for every process step that hasn't been proven commercially, about four more months if the feed is raw solids, and more than five more months if the company didn't put R and D, engineering and operations into one accountable team.
Joao: Okay, that's a lot of RAND. I was promised stories.
Traci: You were. Let's go to four thousand metres.
Joao: Kumtor.
Traci: Kumtor. Kyrgyzstan, in the Tien Shan mountains, sixty kilometres from the Chinese border. The pit runs into a glacier: they used ground-penetrating radar to find voids in the ice that were tens of metres deep. The mining fleet is derated thirty percent just for altitude. Cameco operated it in a joint venture with the Kyrgyz state company. Construction started in May 1995. Ore went into the mill on the seventeenth of December 1996, exactly on schedule, and the first gold was poured on New Year's Eve.
Joao: Sounds like a Series 1 story so far.
Traci: Then the ore arrived. In the first months of 1997 the mill was getting material from the upper benches of the pit, and recoveries were well below what the feasibility study predicted. Final tailings, the gold left behind, ran at more than two grams a tonne in February and March. The reason turned out to be carbon.
Joao: Preg-robbing.
Traci: Preg-robbing. The host rock is carbonaceous phyllite, and natural carbon in the ore does the same job as the activated carbon you put in the leach tanks: it grabs dissolved gold. Except your carbon gets recovered and theirs goes to the tailings dam. And here's the reference-class twist. The feasibility study samples had come from the old Soviet exploration drifts, which were in the central stockwork zone. The upper benches were more carbonaceous and more oxidised than anything the study had tested.
Joao: So the samples weren't wrong, they just weren't the ore you mine first.
Traci: Exactly. They were representative of the middle of the deposit, and the plant was being fed from the top. So what did the team do? Some of it is classic metallurgy: masking the natural carbon with diesel and a surfactant, selectively stockpiling the worst ore, regrinding finer, nearly tripling the carbon concentration in the first leach tank. And some of it is lovely improvisation. They noticed that red food dye sticks to carbonaceous material the same way gold does, so they built a thirty-minute operator test around food colouring, which let the control room see trouble coming in the mill feed.
Joao: Food dye. That's brilliant.
Traci: It is. There's also a very unglamorous problem in there. The local quicklime was sometimes as low as fifty percent active, against eighty-plus for typical North American lime, the acid for washing the carbon couldn't be delivered in enough quantity until September, and lime scale was coating the carbon. Nobody writes that into a feasibility study.
Joao: And how did the year end?
Traci: By the last quarter the mill was running slightly above its design throughput, at its budgeted recovery. And for the full year, Kumtor produced about five hundred and two thousand ounces against a budget of about four hundred and thirty-five thousand.
Joao: Hold on. They had a recovery crisis for half the year, and they beat budget by fifteen percent?
Traci: They beat budget.
Joao: Then this is my first exhibit. If you can spend six months fighting the ore and still beat your first-year number, the first-year number wasn't a forecast. It was a cushion.
Traci: Or it was a forecast that had already assumed a ramp-up, which is what you're supposed to do. The forecast for 1998 was about six hundred and forty thousand ounces, so the first-year budget was roughly two-thirds of a normal year. That's a feasibility team saying: we'll start in December, we'll have problems, plan for them.
Joao: Which is my point. Somebody decided how much cushion to put in. The curve didn't decide.
Traci: Keep that thought for twenty minutes, because I think it's the real argument of the episode. First I want to do the other end of the distribution, because Kumtor is a story with a happy ending, and the curves exist because of the ones that didn't have one.
Joao: Back to Western Australia.
Traci: Back to the three laterite plants. Cawse, which we started with, Bulong, and Murrin Murrin. The technology at the heart of all three is high-pressure acid leaching. You take a low-grade nickel laterite, mix it with sulphuric acid, and cook it in a titanium-lined autoclave at about two hundred and fifty degrees and forty-five atmospheres. It had been done for decades at Moa Bay in Cuba. So, not new in principle.
Joao: But.
Traci: But every one of these plants changed things. Different autoclave designs, different downstream recovery chemistry, different equipment. Murrin Murrin, the biggest, designed for forty-five thousand tonnes of nickel a year, used equipment that departed from the Moa Bay precedent, for good reasons, to save energy and cost. It produced practically nothing in its first year. Three and a half years after start-up it was at about seventy percent of design. McNulty's 2014 update puts both Moa itself and Murrin Murrin below his Series 4 line.
Joao: And Cawse was the best of the three?
Traci: Cawse was the best, and the paper its operators wrote is one of the most honest things in the pile. It's called Commissioning Under a Microscope, which tells you how watched they felt. A couple of details. The titanium block valves in the autoclave circuit were a type proven in gold pressure-oxidation plants. Under nickel laterite conditions their life was several days. The team redesigned them in-house and got the life to several months.
Joao: Days to months is quite a design change.
Traci: And here's the one that I think is the most instructive. The autoclave itself was extracting more than ninety percent of the nickel from almost the first day. But overall recovery, the nickel that actually came out the end as metal, ran at sixty-five to seventy-five percent for a year. The problem wasn't the exotic part at all. It was the washing circuit, the counter-current decantation thickeners, which are about as conventional as hydrometallurgy gets. Once they got the thickener underflow density under control, recovery jumped. Design recovery for nickel came in March 2000.
Joao: So the new technology works, and you lose a year on the boring bit.
Traci: Which is McNulty's line from 1998, almost word for word: the innovative sections may have worked well, but simple operations hadn't been carefully engineered. The grinding circuit at Cawse was designed for hard rock, and the ore turned out to be full of clay, during record rainfall. They ended up renting screens to bypass part of the crushing plant.
Joao: You said there was a reference-class detail in there too.
Traci: There's my favourite one in the entire episode. The feasibility study estimated a workforce of one hundred and thirty-five. The operators explain where that number came from: it was based on a standard gold plant model, because there wasn't a comparable project to compare with.
Joao: Oh, that's perfect. They needed an outside view, couldn't find one, and borrowed the wrong reference class.
Traci: They borrowed the nearest reference class to hand, which was a gold mill, and a pressure acid leach and solvent extraction refinery is a chemical plant as much as a mineral-processing plant. They recruited one hundred and fifty-three to start with. Six months in, staff turnover was damaging the operation, and they added another twenty-six positions.
Joao: And then the money.
Traci: Then the money. Rob Nice, an independent engineer, wrote this up in 2003. Capital costs for the three plants came in seventeen to twenty-eight percent over their feasibility study estimates. Bulong was sold to a junior company just as commissioning began, and was later taken over by its bondholders and put into receivership. Centaur, the Cawse owner, went into receivership in March 2001. Murrin Murrin's owner had to raise a large rights issue to pay out its bondholders. Nice's phrase for the pattern is that the second or third owner is the one that reaps the rewards.
Joao: And this is where ramp-up stops being an engineering question.
Traci: Exactly where. Dynatec's Finlay Campbell and colleagues, the Sherritt engineers who'd designed laterite plants for decades, ran the numbers in 2004 for a hypothetical laterite project. Over the first three years, the difference between a Series 1 and a Series 3 ramp-up was worth about twenty percent of the project's capital cost. A Series 4 project, in their example, finished year three with more than a hundred million dollars of operating losses piled on top of unpaid capital.
Joao: Twenty percent of capex is basically the contingency, gone, from the ramp-up alone.
Traci: Before you've spent a cent fixing anything. And Campbell makes a point, which I'll paraphrase, that explains why bad ramp-ups stay bad. If a plant spends a year running at thirty or forty percent because of one deficiency, the bottlenecks that would hold it at seventy, eighty or ninety percent haven't even been found yet.
Joao: Because you never got far enough up the hill to see them.
Traci: You only find the next constraint by getting past the current one. And while you're stuck, the cash is burning and the bond coupons are due.
Joao: Okay. I've been patient. Can I make my case now?
Traci: Go.
Joao: Right. Here's my claim. The McNulty curve has become a kind of law of nature in this industry. Lenders use it, independent engineers use it, Hatch says it's widely used by operating companies and project financiers to see whether a project is ahead of or behind. And I think once a curve becomes the thing you're measured against, it stops being a neutral description. People plan to the curve, they put enough cushion in to beat it, and then they publish the paper about beating it.
Traci: Okay. Evidence.
Joao: Exhibit one, Century. Pasminco's zinc mine in northwest Queensland, at the time the largest zinc mine in the world. Very fine-grained ore, difficult mineralogy, ultrafine regrinding to about six microns. Hard project. The concentrator starts in November 1999. The 1996 feasibility study had a ramp-up learning curve built in, and the paper written by the metallurgy team plots actual production against it. The study curve has them reaching about two-thirds of nameplate within half a year, mid-eighties after a year and a half, and the full five hundred thousand tonnes of zinc a year only after about three and a half years.
Traci: That's a slow curve.
Joao: It's a very slow curve. And the plant was ahead of it nearly every month. They first hit nameplate rate in January 2002 and sustained it for three months running by August 2002. Their paper says production was consistently ahead of the learning curve. Well, of course it was. The curve was written to be beaten.
Traci: Or it was written to be realistic, and they were good. Mackey and Nesset looked at Century in 2003 and classed it as Series 2, which is about what a fine-grained, carbon-bearing, ultrafine-grinding project ought to expect.
Joao: Exhibit two, Cerro Verde 2. Freeport's second concentrator at Cerro Verde, near Arequipa in Peru. Two hundred and forty thousand tonnes a day, which at the time was the biggest concentrator ever built in one go. Start-up in September 2015. Their paper plots a McNulty-type benchmark curve, a composite of about thirty similar concentrator ramp-ups, which says nameplate on the first of August 2016, eleven months in. The plant hits nameplate on a thirty-day average on the nineteenth of February 2016. Day one hundred and fifty-six.
Traci: Five months.
Joao: Five months. Six months ahead of the benchmark. So either Freeport's engineers are wizards, or the benchmark is too slow to mean anything for a modern, well-run project.
Traci: Can I answer that one now? Because I think it's the key to the whole argument.
Joao: Go on.
Traci: Look at what Cerro Verde 2 actually was. It copied the process of the existing concentrator at the same site. The paper says fifty percent of the experienced staff from the old concentrator transferred to the new one. Operators had four months of training on the existing plant before start-up. The operational-readiness programme started eighteen months before first ore. And it was commissioned in phases, one ball mill at a time, so the crews learned on two mills before running six.
Joao: So it's a copy, next door, with half the old crew.
Traci: It's practically the Series 1 checklist made concrete. Mature technology, equipment like the last lot, a team that has run it, ore that's well understood. The curve didn't fail at Cerro Verde. The factors that McNulty listed predicted it'd be fast, and it was fast. The thirty-plant composite is a reference class for concentrators in general. Cerro Verde 2 belonged in a narrower and better one: copies of a plant you're already running.
Joao: Fine, that's a fair hit on Cerro Verde. Exhibit three is Constancia, and I think this one's harder for you. Hudbay's copper mine in the Peruvian Andes. Design throughput seventy-six thousand tonnes a day. First concentrate December 2014. They run at design for three days straight at the end of March 2015. About three months.
Traci: That's quick.
Joao: Read the design section. Ausenco, the engineer, worked out during optimisation that because the first five years of ore are softer supergene material, the grinding circuit could treat up to eighty-six thousand tonnes a day in that period. The nameplate of seventy-six thousand is really a number for the harder ore that comes later. And by the time the paper was written, they were running at ninety-five thousand by letting the grind go coarser.
Traci: So the plant was designed for harder ore, and it got soft ore first.
Joao: And so a fast ramp-up to nameplate tells you something about the nameplate, not just about the ramp-up. That's my sandbagging point. If I want to look like Series 1, I set nameplate on the hard ore, start on the soft ore, and I'm a genius in three months.
Traci: I don't think that's quite fair to Hudbay, because the hard-ore basis is a legitimate design choice for a twenty-year mine. But as a reviewer's point, I agree with you completely. When someone shows you a fast ramp-up, the first question is: nameplate on what ore, and was there margin in the design? Constancia actually shows the opposite problem too. Throughput beat forecast, but recovery lagged, because the early ore had more oxide copper than expected and flotation can't recover oxide.
Joao: And my last exhibit isn't a mine, it's a definition. What does "commercial production" mean?
Traci: Ah. Yes.
Joao: New Afton, in British Columbia, 2012. Their target for declaring commercial production was sixty percent of design rate. They got there about two weeks after start-up. Côté Gold, in Ontario, 2024: commercial production defined as thirty consecutive days at an average of sixty percent of nameplate. Carmen de Andacollo in Chile, 2010: thirty days at seventy-five percent of design, plus forty-eight hours at a hundred. Las Chispas in Mexico, 2022: two months at eighty percent. These are four perfectly normal definitions and they're all different.
Traci: And none of them is nameplate.
Joao: None of them is nameplate. So when a press release says "commercial production, ahead of schedule", an investor hears "it works", and what it actually means is that it's run at sixty percent for a month. And the McNulty curve itself, which you're holding up as the outside view, has its own definitional wobble. You told me about it earlier.
Traci: I did. In the 1998 paper, where you couldn't separate commissioning from start-up, time zero was the end of construction. In the 2004 paper, McNulty restates it as six months after the end of construction. And the 2004 paper also says the average Series 1 plant took eighteen months to reach design, which doesn't match the 1998 chart, where Series 1 is essentially at design by twelve. Small things, but they matter if you're using the curve to the month.
Joao: And one more, then I'll stop. In 2014 McNulty publishes a new, faster curve, Series 1A, and says owners shouldn't be too pessimistic about early cash flow if they've done their homework. It's built from the best twenty-three plants out of eighty-one examined. And for one of the data sets, Freeport's plants, he says explicitly that only those better than the Series 1 average were included, because he didn't know why the others had underperformed.
Traci: So Series 1A is the winners' curve by construction.
Joao: By construction. Which is fine as a picture of what's possible. It's not a reference class. If I put Series 1A in a bank model, I'm doing exactly the inside-view thing Flyvbjerg warns about, just with a nicer chart.
Traci: Okay. That was a good case, and some of it I'm going to concede. But let me push back on the core claim, that the curve is mostly a budgeting habit.
Joao: Please.
Traci: If it were just a budgeting habit, you'd expect projects that ignore it to do fine. They don't. Take Red Chris, Imperial Metals' copper-gold mine in northern British Columbia. Conventional porphyry concentrator, thirty thousand tonnes a day. We have the 2012 technical report, and its production schedule has the mill processing the full eleven million tonnes in year one. No ramp-up at all.
Joao: Full rate from day one.
Traci: Full rate. The mill started in February 2015. Imperial's own production releases say it averaged about twenty-five thousand seven hundred tonnes a day for the rest of 2015, and about twenty-six thousand four hundred in 2016. So roughly eighty-six percent in the first year and eighty-eight in the second, against a plan that said a hundred.
Joao: Fair. That's a curve-ignorer getting caught.
Traci: And it's not unusual. Mount Milligan, also in British Columbia, a conventional SAG mill and flotation plant, the most standard flowsheet you can have, started in September 2013 and reached its sixty-thousand-tonne nameplate in December 2015. More than two years, with pre-crushing added to get there. RAND's database shows planned start-up times that tell you almost nothing about actual ones. Merrow's half of plants missing eighty-five percent in months seven to twelve. And here's one more that I think goes right at your point. McCarthy, reporting in 2003 on a study he'd done with Ward, looked at nine Australian underground base metal mines. Most of them could run for three months at fifteen percent above design.
Joao: Above.
Traci: Above. And over the longer term they averaged only about ninety-two percent of design. Only half reached design throughput by year three, and a quarter never did. There's a honeymoon. A plant can sprint. Whether it can sustain is a different question, and Hatch's newest post, Part 9, is about exactly that: projects that look like Series 1 for a few months, then drift down to a lower curve once commissioning teams leave and operations inherits the problems.
Joao: So the sprint looks like proof, and it isn't.
Traci: It isn't. And the Hatch posts say that, in the McNulty update, more than a hundred and fifty projects have now been tracked, and they talk about why a quarter of ramp-ups still fail today. Again, that's the blog's summary; we haven't seen the underlying data. But nothing in the modern evidence says ramp-ups got easy.
Joao: So where does that leave us? Is it a law of nature or isn't it?
Traci: Here's where I've ended up, and tell me if you can live with it. It's not a law of nature. It's closer to a law of organisations and of solids. The shape is real: new plants take time, and solids plants, novel plants and inexperienced teams take longer. The exact curve isn't sacred, and you're right that it mixes definitions and that the fast version is cherry-picked. So you use the curves as the prior and the factor list as the adjustment, which is exactly the Flyvbjerg recipe: anchor on the reference class, then adjust for what's genuinely different, and make the adjustment earn its keep.
Joao: I can live with that, with one amendment. Before you anchor on anything, you read three definitions: what nameplate is based on, when time zero is, and what "commercial production" means in this particular press release. Half the apparent miracles disappear when you do that.
Traci: Agreed. And the other half are usually Cerro Verde: a plant whose real reference class is narrower and friendlier than "concentrators".
Joao: Okay. So the factor list. What actually moves a project from Series 1 to Series 3?
Traci: Five things in the brief, and the literature lines up on all five. Let's go fast, with one example each. First, novel technology, but measured properly: not "is this new?", it's "how many steps are new, and how much of the design rests on data from a real plant?" RAND's ten or eleven weeks per unproven step is the cleanest rule. And the cautionary tale is the magnesium project.
Joao: Which one?
Traci: Australian Magnesium Corporation. A new, patented process for magnesium metal. They did everything right technically: two pilot plants and a demonstration plant over eleven years, around two hundred million dollars of testing and demonstration, independent cold-eye reviews. Terry McNulty himself reviewed their ramp-up and rated it Series 1.
Joao: And?
Traci: And it never started up. The majority owners changed during execution, their main businesses weren't magnesium, the project became dependent on project finance, and the lead engineering contract had to be converted to a lump sum part way through engineering. The contractor priced the risk high, the two sides couldn't agree, and the project was abandoned. It's in the 2011 Hatch paper by Wasmund and colleagues. The curve rates technical readiness. It can't rate whether the capital structure survives.
Joao: Which is Cawse again, really.
Traci: It's Cawse from the other side. Second factor, fast-tracking. RAND found that for first-of-a-kind plants, planning eight months or more of overlap between detailed engineering and construction typically added more than half a year to the actual schedule, because construction ends up waiting for engineering to catch up. Tenke Fungurume in the Congo is the interesting counterexample. Freeport deliberately started up seven months early by running only part of the plant and buying acid until their own acid plant, a prototype, was ready. They hit design copper and cobalt in months six and seven. Fast-tracking can work when it's designed in as a phased start, rather than squeezed into the engineering.
Joao: Third?
Traci: Feed variability, and here the geologists in the audience will nod, because it's the one the plant can't fix. Kumtor is the story: the ore you test isn't the ore you mine first. Ernest Henry in Queensland is the quieter version. They decided at feasibility not to run a pilot plant for ore variability and to rely on bench tests. Mill throughput came up like a Series 1 plant. Copper production lagged like Series 2, because of ore variability. Throughput and recovery ramp up on different curves.
Joao: Fourth, operational readiness.
Traci: Tropicana in Western Australia is the most candid. Their mining team measured how long recruitment really took: fourteen weeks from advert to someone starting, and then about three more months before they were fully effective. Their engineering and planning group didn't expect to be fully effective until November 2013, which was four months after the mine was at full production and seventeen months after mining started. So the people who're meant to plan the mine were still arriving while it ramped up. On the money side, Gary Tostengard argued back in 1988 that half a percent to one percent of plant capital spent on start-up preparation was the difference between a fair start-up and a good one. And a 2003 feasibility-review checklist by Shillabeer and Gypton tells reviewers to expect staff turnover of thirty to fifty percent in the first six months.
Joao: And fifth, owner versus contractor.
Traci: This one has the clearest natural experiment. Back to the laterite trio. Murrin Murrin was built on a lump-sum, fixed-price contract. Nice reports that no Murrin Murrin technical people were in the engineer's offices during design, part of the process wasn't finalised when the contract was awarded, and commissioning started about seven months late. Cawse used a reimbursable contract with a strong, experienced owner's project manager, and it was the only one of the three to reach practical completion on the contracted date, at a final cost below its definitive estimate.
Joao: So the curve isn't just about the plant, it's about who's in the room.
Traci: Which is the whole of Hatch's Part 7, called Accountability Without a Home. The people who make the decisions that set the ramp-up, in the feasibility study and early engineering, are usually gone by the time it happens. Responsibility is assigned at delivery, authority sits years earlier. And the Wasmund paper has the best line about fixed-price contracts on new technology: they lead to arguing over who pays for the modifications that are inevitable after start-up, when a day of earlier production is worth more than settling the exact split.
Joao: Okay. Test cases. You promised we'd take the curves to real technical reports.
Traci: Four of them, each a slightly different lesson. The first is Côté Gold in Ontario, IAMGOLD's open pit with a high-pressure grinding roll circuit, thirty-six thousand tonnes a day. And we have two versions of the study, which is the fun part. The 2018 feasibility study says the plant is expected to take ten months to reach design throughput.
Joao: Ten months. That's a Series 1 assumption.
Traci: It's a Series 1 assumption. Then the 2022 technical report update, same project, same boilerplate paragraph about high-pressure grinding roll circuits taking anywhere from six months to a year, comes to a different conclusion: twenty months to reach design throughput, with ninety percent reached after ten.
Joao: Same evidence, conclusion doubled.
Traci: Same evidence, conclusion doubled. And now we know what happened, from IAMGOLD's own releases. First gold on the thirty-first of March 2024. Commercial production declared in August, on that sixty-percent definition. An average of ninety percent of nameplate for the month of March 2025, about a year in. And the full thirty-six thousand tonnes a day, averaged over thirty consecutive days, on the twenty-first of June 2025.
Joao: So about fifteen months.
Traci: About fifteen months. Slower than the 2018 study, faster than the 2022 one, and about two months behind the 2022 study's ninety-percent milestone. The Hatch Part 1 post actually cites a 2025 paper on Côté as an example of how the curves are used. We haven't read that paper. But the arithmetic from the company's own announcements is clear: the inside view moved towards the outside view between 2018 and 2022, and reality landed between them.
Joao: Second?
Traci: Red Chris, which we've done. Full rate in year one on paper, about eighty-six percent in reality. Third is Las Chispas, SilverCrest's silver-gold mine in Sonora. We looked at its feasibility study last episode for the equipment numbers. On ramp-up it's the good example. The study explicitly built a stockpile so the plant would have ore while the narrow-vein underground mine ramped up gradually, out to 2025. And SilverCrest's release from January 2023 compares the two. Plant throughput in the fourth quarter of 2022 averaged eleven hundred and thirty-five tonnes a day, against a feasibility plan of eleven hundred and forty-three. December beat nameplate. The mine, meanwhile, ran at seven hundred tonnes a day against original guidance of seven hundred and fifty, and the stockpile was smaller than the study had expected.
Joao: So the plant was on plan, the mine was a bit behind, and the stockpile absorbed the difference.
Traci: Exactly. Decoupling the mine from the mill turned a mining ramp-up problem into a stockpile number, which is what a buffer is for. And the fourth case is the one with the opposite shape. Platreef, Ivanhoe's platinum-group-metals mine in South Africa, which was also in our last episode. The 2025 feasibility study has the first, small concentrator, seven hundred and seventy thousand tonnes a year, starting in the fourth quarter of 2025, and the mine stepping up to four million tonnes a year by 2029 as new shafts come in.
Joao: And did the plant start?
Traci: On time. First concentrate in November 2025. But the plant was waiting on the mine. Before Shaft 3 came in, hoisting capacity was the bottleneck, so the concentrator was run in campaigns on lower-grade development ore. Ivanhoe's own second-quarter 2026 report says the phase one concentrator milled about thirty-seven thousand tonnes in the quarter. Against a plant sized for seven hundred and seventy thousand tonnes a year, that's about a fifth of its quarterly capacity, by our arithmetic.
Joao: A fifth. And the shaft?
Traci: Shaft 3 was completed on schedule. Ivanhoe now expects commercial production in the fourth quarter of 2026, where earlier in the year they'd said mid-year. And in fairness, this was a phased plan that always expected the mine to take years. But it's a textbook case of something an old feasibility manual from Kilborn and Micon put plainly in 1991: simple mills usually reach full capacity in their first year, and mines, especially underground mines, often take longer to catch up with the mill.
Joao: So the ramp-up curve you care about isn't the plant's. It's whichever one is slower.
Traci: It's the slower of the mine, the plant and the people. And it's usually not the plant.
Joao: Right. Let's land this. Rules of thumb. I want things a reviewer could actually write in the margin of a feasibility study.
Traci: Ten. They're in the show notes with sources and the conditions under which they apply, so here's the spoken version.
Joao: Go.
Traci: One. A greenfield conventional concentrator should plan on roughly a year to sustained nameplate, and a first-year average of something like seventy-five to ninety percent. A schedule at a hundred percent in year one is an inside view. Red Chris planned it and got eighty-six.
Joao: Two.
Traci: Two. For a novel flowsheet, a hydrometallurgical plant, or a pioneer plant fed with raw solids, start from a Series 3 profile, about sixty percent of design in year one and around eighty in year three, and make the project argue its way up. RAND's ten or eleven weeks per unproven process step is the adjustment. Remember none of the eleven solids pioneers met their schedule.
Joao: Three.
Traci: Three. Price the ramp-up as money, not as a date. In Campbell's laterite example, a Series 2 ramp-up instead of Series 1 cost about ten percent of capital over three years, and Series 3 about twenty. RAND's first-of-a-kind plants spent about nine percent of capital on start-up itself. If the contingency is ten percent and the ramp-up is novel, the contingency is spoken for.
Joao: Four is mine.
Traci: Four is yours.
Joao: Read the definitions before you read the curve. What is nameplate based on, which ore, which grind, how much design margin? When is time zero? And what does "commercial production" mean, because it's anywhere from sixty to eighty percent in the cases we read. A fast ramp-up to a soft nameplate isn't evidence of anything.
Traci: Five. Ask which ramp-up is the slow one. Underground mines usually lag the mill. A planned stockpile buffer, like Las Chispas, is a good sign. A plant that depends on a shaft that isn't finished, like early Platreef, will run at whatever the hoist allows.
Joao: Six.
Traci: Six. Discount the honeymoon. Three months above design isn't proof of sustained capacity. McCarthy's mines ran fifteen percent over for a quarter and averaged ninety-two percent over the long term. Ask for the thirty- or ninety-day average, and ask again a year later.
Joao: Seven.
Traci: Seven. Look for the people. Is operational readiness funded, something like half a percent to one percent of plant capital, and does it start a year and a half before first ore, as at Cerro Verde 2? Are experienced operators transferring from a similar plant? Is turnover of thirty to fifty percent in the first six months in the plan? Does the workforce estimate come from a genuine analogue, or from the wrong one, like Cawse's gold-plant model?
Joao: Eight.
Traci: Eight. Look at the contract and the owner. A lump-sum, fixed-price contract on new technology with no owner engineers in the design office is a Series 3 or 4 flag. A strong owner's team with operators in it is a Series 1 or 2 flag.
Joao: Nine.
Traci: Nine. Make sure the balance sheet can survive the slow case, not just the base case. Cawse reached design and the owner still went into receivership. Australian Magnesium was rated Series 1 and never started. Rob Nice's advice was to have a separate facility for overruns. Ramp-up risk is financing risk.
Joao: And ten.
Traci: Ten. Use the curves as the prior, never as the forecast. Pick the narrowest honest reference class: copies of a plant you already run, conventional concentrators, novel hydromet plants. Anchor on its typical outcome, then adjust with the factor list, and write down why. Don't use the best-performer Series 1A curve as a base case.
Joao: If someone only remembers three of those?
Traci: For me it's two, nine and ten. Solids pioneers take longer than anyone plans. The balance sheet has to survive the slow case. And the curve is the prior, not the answer.
Joao: Mine are four, five and six, because they're the cheapest checks. Read the definitions, find the slowest ramp-up, and don't believe the honeymoon. You can do all three with a technical report and a stack of press releases in an afternoon.
Traci: Which is a good place to stop. One last picture, from Hatch's Part 6. In 1916, the chemist Leo Baekeland, the inventor of Bakelite, gave a medal lecture and said, roughly, that you should make your mistakes in the small plant and your money in the big one. Every story in this episode is a version of that. Falcondo, in the Dominican Republic, ran two integrated pilot plants at the mine site for years, moved more than ten of the pilot crew into the owner's team, and went from fifty-nine percent of design in its first year to essentially a hundred in its second, with a completely new process.
Joao: And the laterite trio made their mistakes in the big plant.
Traci: In the big plant, in public, with the bond coupons running. The ramp-up curve is just the record of where people chose to make their mistakes.
Joao: That's the show. The full rules of thumb, the test-case comparisons and every source are in the show notes. We'll be back with the regular episodes.