Monthly Market Update – Sep 2026
Performance data and commentary across asset classes and strategies.
Tables are sorted by 3-month return for the quarter end. Benchmark rows are highlighted.
Data current as of end-Sep 2026.
What’s New
- Mortgage ratesEvery Australian lender’s current home loan rates.
- Term depositsRates by term, amount and bank.
- Savings accountsBase and bonus rates at any balance.
- Super LabCompare performance, fees, asset allocation and default insurance across any industry super fund.
Fixed Income
- Absolutely brutal month for anything with any kind of duration.
- The only protection was being as short duration as possible.
- Less than half the funds here have beaten CPI over 12 months.
- Credit Funds with negative 1Y returns look interesting, just the balance between duration and spreads.
- The question is how high can rates get?
- For the US10Y, is this Oct-2021 (1.50%) or Oct-2023 (5.01%)?
- I ask you the question, have bond yields made a local peak?
| Fixed Income | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa |
|---|
Monthly update, straight to your inbox.
Private Credit
- Sector is still cooking.
- I am comfortable in saying the worst of it hasn't even arrived yet.
- Give us another rate rise RBA, will do a world of good for the CRE sector.
- Remember, be opportunistic! You don't get paid to be a bear!
| Private Credit | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa |
|---|
Practice automation
Client Portal
How do you interact with your clients today? Do they get a PDF or can they access your portal and update their details while reviewing their own files automatically?
Domestic Large Cap
- Holy moly, Bennelong the only one with a positive month! What?!?
- Utterly shocking 5 years for active managers with less than 10% outperforming the ASX200.
- You genuinely couldn't be a long term fundamental investor in this market unless you did something different.
| Domestic Large Cap | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa |
|---|
Domestic Mid/Small Cap
- I genuinely worry about this part of the market.
- Lots of funds with outflows and many shutting shop.
- Means capital exiting along with the CGT discount tax changes means the sector is full of headwinds.
- Noting that there was excess capital that ended up in midcaps as many investors (wealth and insto) that ended up in the space trying to find "growth" outside the banks and commodities.
| Domestic Mid/Small Cap | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa |
|---|
Monthly update, straight to your inbox.
Domestic Micro Cap
- Rarely do you see a fund almost halve in 12 months, yet here we are. Ouch...
- A larger proportion of names here have outperformed large caps.
| Domestic Micro Cap | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa |
|---|
International Equities
- Growth snapped back last quarter but gave most of it up in October.
- MSCI World +19%pa and +13.3%pa over 3yrs & 5yrs respectively, this is not normal.
- Tread lightly, past performance is not an indicator of future performance.
- Exposure to big name brands with long track records has been a painful experience this month.
- Bottom of the table for the quarter full of Global small caps.
| International Equities | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa |
|---|
Infra + REITs
- Utterly filthy month and quarter for infrastructure, all about the US10YR.
- Whenever you see a drawdown of this quantum combined with the notion that AI growth isn't slowing and utilities being underwritten by reshoring, electrification, and AI, what do forward returns look like from here?
- Asking again, is this Oct-2021 or Oct-2023?
| Infra + REITs | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa |
|---|
Monthly update, straight to your inbox.
Other
- Huge Quarter for GDX but a big negative month.
- The trend funds have absolutely killed it last quarter and the last year, check out AQR/Aspect/GCM/Winton.
- Have to remember that their aim is to generate <Cash + spread> with that spread being circa +300-500bps.
- Your historical bias of thinking the trend funds cannot generate good returns is no longer true under that thinking, rates are no longer falling.
- Though when trend has a poor month while markets are ripping makes a terrible client conversation.
- Gold eases to a +1.8% return over the year and inflation beats returns.
| Other | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa |
|---|
Monthly update, straight to your inbox.
GICS Sectors
- Discretionary, REITs, Tech and Telecoms slammed over 3 months (and 1yr), main street in pain.
- I wonder where "growth" comes from in our index. Lots are chasing and looking for this when thinking about allocating.
- But the more I think about it, we have mature "value" businesses, and that is OK!
- The ASX200 doesn't need to be "growth", it is allowed to be a more value biased index that is focused on generating yield.
- More an allocation decision for you dear reader than anything else.
| GICS Sectors | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa | 10 Yr pa |
|---|
Crypto
- A big big quarter for this space.
- I point to this article for reference.
- The lore is crypto markets set a bottom in October.
| Crypto | 1 Mth | 3 Mth | 1 Yr pa | 3 Yr pa | 5 Yr pa |
|---|
Monthly update, straight to your inbox.
Thoughts of the Month
Firmus, my AI bill, and portfolios for non-average times
A lot of AI talk this month (again, sorry) but I think it creates non-negotiables and is a source of volatility in portfolios.
Firmus
Firmus pulled its IPO after being marketed at A$11 a share, then A$9, then A$8.25, then A$5.50, and the whole thing went back to "private markets".
The IPO makes me wonder whether this was an IQ test for those trying to raise capital or those thinking about allocating capital.
I'm leaning towards the bankers being IQ tested, because this was one of the more disgusting pieces of corporate exuberance I've seen in a while.
Well let's use this opportunity to review the Firmus IPO from a basic, bottom up perspective and also cover off valuations.
How do you value a datacentre?
As Aussies, and I feel insane saying this, but we have an acute sense of how much power costs as securing an abundance of cheap power in Australia is challenging.
At the very crux of things, the whole datacentre/AI/neocloud trade is to turn watts into tokens. Nvidia have nailed it and benchmark things to token throughput per megawatt.
The demand for tokens (a commodity) well and truly outstrips the supply (power). Yes there are other inputs (datacentres, GPUs etc) but you can build everything relatively quickly in comparison.
This is the literal AI trade from first principles driving the global markets today, from infra/utility assets through to semiconductors.
If you don't have power, you don't have anything. Amazon leaked that they have GPUs sitting there unable to be deployed because they don't have enough power secured.
You can plug in the same Nvidia GB300 GPU anywhere in the world, be it in a Google datacentre or a Firmus one, and it earns the same amount. As long as you have a stable, cheap, reliable source of power.
Watts = Tokens.
(Tokens = Knowledge work).
A Bird in the Hand...
You can split every AI cloud company into 2 buckets:
- "Actual Watts" - Power that is live today. I will value this at what it costs to build.
- "Potential Watts" - Power secured but not yet built. Take the EV, add any CAPEX needed to build, subtract the "Actual Watts" bucket. The leftover value is what you're paying for the future.
If we know what the value of each actual watt is, then we can infer the value of each potential watt.
And then we can ask the most important question of all!
How much are you paying for each Potential Watt above what it costs to build?
If you're paying above market value of what an "Actual Watt" is today, this is what I like to call my "Hope Premium".
And remember kids, hope is not a strategy!
Why don't we put this in a neat little table:
Fully-loaded EV split by live vs secured-unbuilt power, US$, 9 October 2026 close
| CoreWeave | Nebius | Firmus @ A$11 | Firmus @ A$9 | |
|---|---|---|---|---|
| Market cap | $45B | $60B | $31B | $25B |
| Fully-loaded EV | $162B | $213B | $61B | $55B |
| Actual Watts | ||||
| Live power | 1.85 GW | 0.9 GW | 0.046 GW | 0.046 GW |
| Visible value | $68B | $33B | $2B | $2B |
| Potential Watts | ||||
| Secured, unbuilt | 2.35 GW | 4.1 GW | 0.865 GW | 0.865 GW |
| Implied future value | $94B | $179B | $59B | $53B |
| % of value in the future | 58% | 84% | 97% | 96% |
| Value per potential GW | $40B | $44B | $68B | $61B |
| Build cost per GW | ~$37B | ~$37B | ~$42B | ~$42B |
| Hope premium per potential GW | $3B | $7B | $26B | $20B |
Fully-loaded EV = EV + capex to build all secured power. CoreWeave and Nebius are costed at US$37B per GW, mid of Bernstein Oct esti. CRWV live power is its CY2026 guidance (>1.85GW) vs 4.2GW contracted at its Aug results. NBIS is its CY2026 target of 0.8–1GW connected and 5GW contracted. Firmus figures are from the offer documents as reported: 46MW operating, 865MW in development at a capital cost of A$51.7B, funded by ~US$30B of forecast debt plus the IPO cash. Firmus @ A$9 was the reported price cut. Market caps use basic share counts, A$1 = US$0.70. If you're still reading - the build cost of the future GW washes out, so the hope premium is just today's EV, less the cost of what's already live, divided by the GW still to build.
So when this deal died, Firmus wanted 3x what the market was paying Nebius for a future GW, and nearly seven times CoreWeave.
New investors were being asked to pay a premium for "Potential Watts" in a rising rate and rising inflation environment where all input costs are going up and margins are falling.
Egregious.
Firmus.
Only 5% of their total wattage is "Actual Watts" today. So at A$11, 97% of the value was ascribed to pay for power IN INDONESIA.
A country that just got kicked out of MSCI, known for their stable, cheap, reliable power, secure water sources, and their ability to deliver infrastructure on time and on budget.
The last raise was done at a third of the price just two months prior, and the majority (55%) of the pre-IPO capital could trade on day 1.
3x in 2 months, piss off.
Where could the Hope premium even come from?
A GW costs similar levels to build globally and simply earns the going rate (remember, tokens are commodities!). But maybe, just maybe, Firmus could earn a premium and a Firmus GW could be worth more than anyone else's!
- Price - Charge more per GW? Nope, running the same Nvidia chips as everyone else.
- Cost - Spend less on opex/capex? Except power is a small slice of the cost next to the GPUs and their own forecast build cost per GW sits above Bernstein's range.
- Speed - Nope, they have 46MW running today vs GW at their peers.
- Cost of capital - Nope. Blackstone's opportunistic and credit money is funding them, while the hyperscalers fund the same chips out of free cash flow and IG Bonds.
- Software - Nope, the value is in training your own model, Firmus is a landlord.
- Scarcity - Maybe? I mean how many other AI neocloud names can you buy on the ASX?
Market Blowup
Unfortunately, the neocloud market blew up shortly after the IPO went under. CoreWeave and Nebius are down 25% since the IPO was pulled.
A nice little table on the price based on the "Hope Premium" being charged at the peers today:
Hope premium per potential GW (US$) and Firmus price implied at the peers' premium
| 7 Oct | 9 Oct | 16 Oct | |
|---|---|---|---|
| CoreWeave | $4B | $3B | –$2B |
| Nebius | $8B | $7B | $3B |
| Firmus @ A$11 (hypothetical) | – | $26B | $26B |
| Firmus price at peers' premium | – | ~A$3.80–5.00 | ~A$2.30–3.90 |
This assumes the GW live today have not changed in value. Only the future value of each delivered GW has moved.
But sometimes there is opportunity here. CoreWeave's EV went below the cost to build the 1.85GW it expects to have running before Xmas, which is an extremely interesting scenario and a reminder that being a bear doesn't pay.
The greater fool they are not
Someone does need to do the buildout of this stuff.
The hyperscalers have seen the light and the economics for them are phenomenal. But they've used/committed all their free cash flow, so the big question I'm falling over at is whether growth capital has been exhausted and we're at saturation point.
And why would you ever imagine that equity holders are going to get a cent of margin here in an incredibly capital intense market? The lenders take the margin. The challenge for every one of these neoclouds is funding the buildout without needing to raise further (extremely) dilutive equity, with cash balances bleeding dry on the timing gap between interest costs and enough revenue coming through the door. The higher cost of funding does not help.
Firmus is a signal in this market. Yes, everyone involved tried to take the market for a ride, but underneath I worry, because if this was a few months earlier it would have been an amazing IPO.
The investors who are committing capital to the buildout now are credit and infra allocators who are a tiny bit more discerning. They're not as red-pilled on AI, nor do they have the same convexity that early stage investors have had.
Either the funding is going to dry up or the economics are going to need to change.
Let's all try to do a little bit better.
The market will pay up for AI compute it can see (see SpaceX) and very little for what lies in the future.
At least this month. Who knows, maybe the neoclouds all bounce back?
But next time someone pitches you a "platform" with a big number of contracted power, ask two questions. How much power is live, and how much am I paying for the rest?
Monthly update, straight to your inbox.
AI
Why don't we go through my AI usage and give you an idea of my billing, the subsidies I receive, and how I could manage costs to achieve the same output. Reason is if you're operating a business, this is likely prescriptive of what kind of costs you can expect.
How I use AI
- Webchat - I use webchat for maybe 10-15%, especially around knowledge work where I can send queries and work across mobile > desktop > mobile.
- Coding harness - 90% of my usage is non-web chat.
- Claude Code/Codex - 5-10% of time.
- Pi & Oh My Pi - primary depending on use case.
- Workflows are all mostly <Orchestrator + Worker> setup.
- Orchestrator is whereby you get the best model possible (Opus 5.5 currently) to think and strategise.
- Worker (an "agent") is the model that goes off and executes, with the orchestrator writing the prompts and managing workers.
- There can be more than 1 worker, ie multiple workers operating in parallel if a task can be split so that it can be addressed concurrently.
- Workers are generally a cheap and fast model, doesn't have to be the brains.
- Think of this as brains and hands working together. You don't need to spend $$$ on the smartest model on the hands.
What I use
- Firstly, I use AI for literally everything.
- I use Excel as a visual editing tool these days...
- This is my usage both for working, consulting, building stuff for myself and plenty of other bits & bobs.
- Model usage.
- Google/Gemini - general queries on my computer via Google search, just so damn easy. Plus coding.
- OpenAI/ChatGPT - general queries on my mobile. Plus coding.
- Anthropic/Claude - more complex queries and knowledge work. Primary coding model.
- Meta/Muse - was uber cheap, mostly coding. Haven't touched their latest model, Wonder (Full).
- Deepseek/DSV4.1F - cheap and fast, only for coding.
- Trialled lots of others.
- Have run local models too, some I use actively for various things.
- Gemma4 12B is a sensational writing model, no slop. IYKYK.
What I Pay For
- I pay for subscriptions to ChatGPT, Google, Claude.
- All other models I use via API, ie pay for usage.
- At the current rate, monthly average over the last 6 months has been circa US$200 (I quote USD because I pay for the API use in USD).
My Usage in tokens
- This covers my last 6 months.
- I've spent less than US$1.5k on subscriptions and API fees.
- Usage breakdown across Orchestrator + Workers:
| Token type | Tokens |
|---|---|
| Input (ie what you type into a webchat) | ~150M |
| Output (ie what response you get from a webchat) | ~200M |
| Cache (ie the prior responses and files in a chat that get re-read on every prompt) | ~40B |
| Total tokens | ~40.4B |
| Cost at historical list rates | ~$35.7k |
- Call it approximately $35.7k in token usage based on historical list rates.
Pricing ladder and the subsidies
| Same 6 months of usage | Cost |
|---|---|
| List price | $35.7k |
| Competitive price (Gemini4 Argon + Deepseek V4.1F-1017) | $6k |
| What I actually paid | $1.5k |
- $35.7k is what that usage costs at list price. Only enterprises pay list if they're unlucky.
- $6k is what the same work costs at the competitive price.
- Gemini4 Argon as orchestrator, Deepseek V4.1F-1017 as worker, and I firmly believe it would change little in what I produce.
- $1.5k is what I actually paid.
- So am I being subsidised? Against list, enormously. Against the competitive price, about 4x.
- Are these subsidies real? Or are the margins just so absurdly massive? Both.
Would I do something else if my hand was forced?
- I pay somewhere between $150-200 per month depending on usage types and how much excess capacity I need from an API + testing various stuff.
- If I were forced to move solely to an API, like many businesses have, would I change my approach?
- YES!!!
- Gemini4 Argon (orchestrator) + Deepseek V4.1F-1017 (worker). My total cost would drop from the implied $35.7k to $6k.
- Would I pay $6k over 6 months? Absolutely!
What does this mean?
- This is as bad as AI models will ever be.
- This is also as expensive as AI models will ever be.
- Could I automate more? Sure, but I'm hitting the limit of what I can do.
- Therefore I am spending my time on automating other people's businesses and lives.
- Cost for everything is moving towards zero, and I find it extremely challenging to see how the AI Labs (OpenAI/Anthropic) can justify the lofty valuations ($1tr+).
- It's also the whole problem for the neocloud equity in the Firmus piece above. If the token is a commodity heading to zero, the landlord earns the going rate and the going rate goes to the lenders.
- The other players (MSFT/GOOG/META/AMZN) are going to accumulate all the value here as these services are incremental on margins and rapidly expand their TAMs.
You may or may not find the above interesting, but as someone who spends a lot of time living in AI land (some might call me a nerd, for I am merely a power user) and also spending a significant amount of time trying to understand the hardware and infrastructure aspects of the industry, these help me build my mental model.
That mental model is more critical than ever, because understanding even an iota of this is critical for allocators. This is a trend that drives investment and returns across all parts of the equity market, credit markets and infrastructure today.
The number of PMs I've met that invest in AI but don't use it, despite the barrier to entry being so low, is utterly amazing.
My only real takeaway is go download Claude Code or ChatGPT Codex on your computer, ask your webchat to tell you what to do, and go and experiment.
Once the penny drops that an AI is more than just a webchat and can read/write/create/destroy files on your computer better than you (or for that matter any software engineer ever could), it is absolutely eye opening.
I had my eureka moment a year ago now and if you haven't had it yet, I am so excited for you.
Monthly update, straight to your inbox.
Aiming to be Average
Franco Modigliani is my favourite economist.
I don't know if people should have favourite economists, but he is mine because he managed to find something firm and non-negotiable in the world of economics, which is largely pretending vibes can be explained with mathematical models.
He showed you can't bullshit your way through demographics. His trick was to ignore the aggregate and ask what happens to one human over a lifetime. This is life-cycle economics, which won him the Nobel in '85.
Very simply, there are three periods in a life. From 0 to 25 you dissave, from 25 to 65 you save, and from 65 to 85 you dissave. Save for 40 years, dissave for 45.
And the longer we live, the longer we dissave.
The useful bit is that you can start at the micro and scale it up to any country. An aging population and a low birth rate shift the ratio of savers to dissavers, and that ratio drives dependency on everything else - fiscal capacity, labour, human capital, and financial capital. Japan hit this in the 90s and has seemed stuck in time ever since. Migration is the one lever we're pulling to solve this today, and it helps, but migrants age too. It buys time without changing the shape of the curve.
The longest horizon we'll ever have
Total superannuation savings hit $4.77T at 30 June, with $236B of inflows against $148B of outflows for FY26 - $88B net. With SG finally at 12% that contribution gap is about as wide as it is ever going to get. The pool itself will keep growing because returns dwarf contributions, so on the surface nothing looks different. Underneath, the ownership is changing. That outflow number grows every year as people retire and draw down, and dissavers want liquidity and hate volatility. When the largest pool of capital in the country starts wanting those two things, investment horizons invariably shorten.
Today is the longest investment horizon our super system will ever have. It only gets shorter from here.
Which means long-term decision making within the sector gets disrupted exactly when it is needed most.
Non-negotiables
It is needed because the demand side is not optional. There are capital pressures on our economy and these will happen no matter what:
- AI capex - The hyperscalers have committed their free cash flow and are now borrowing. Hard assets such as power infrastructure have long paybacks.
- Electrification - Grid, transmission, storage which are 20-30 year assets.
- Reindustrialisation - Capacity being rebuilt onshore after thirty years of globalisation.
- The military industrial complex restarting, globally.
- Infrastructure - at every level of government to support density.
- Sovereign issuance to pay for all of the above, which has to compete in the same capital markets.
Every one of those has duration over a decade, and the biggest pool of capital in the country (and the same is true offshore) wants it back on a shorter and shorter clock and demands income too. Put that next to the sources of volatility already in play (Brent futures up about 80% YoY, Russia stacking tanks on the border, a middle class that feels it got the short end of globalisation) and something has to give.
And what gives is pretty clear: higher rates, higher inflation, and persistently higher volatility than the long-run assumptions allow for.
Governments then hide it (which is their job) via subsidies & caps to smooth the pain for the average household, but the smoothing spends scarce fiscal capacity to mute the price & inflation signals. So the signal that would tell us to adjust is being deliberately sanded down, and at some point it begins to matter.
The case against
This is just a view, so we need to build an argument against it too.
Rates might mean-revert as they have before, but the trouble is that every long-run assumption brings them back to some benign, arbitrary level while acknowledging that rates might take a volatile path. The volatile path is the bit that kills you.
Migration might carry us, but it's doing a lot of work already, and there's probably a natural ceiling on how much more it can do. It also does nothing to change the non-negotiable macro scenarios.
AI might help here, because I use it all day and I can see the productivity, but the gains are arriving as cheaper tokens, which is deflation in a digital good, and digital deflation doesn't touch the real world and hard assets.
And the long-run average might just be the long-run average. This is the one I keep coming back to. 30/40/50 years of data deserves respect, but the average contains the regime, and thirty years of falling rates is one sample.
I've tried each of these on and none of them changes the super numbers. The horizon is shortening regardless.
What that does to a portfolio
Mean-variance optimisation takes three inputs - returns, vol and correlation - all of which are estimates and predictions (take your pick from where; I see BlackRock Capital Market Assumptions used everywhere). I have no problem with estimates, assumptions and predictions, but they give us fake precision, which makes me deeply uncomfortable.
The practical question is what to do instead, and I'll be honest that this is where it gets hard. Assets that are both asymmetric and uncorrelated with everything else are rare because crash beta kills you every time, and rebalancing into them actively is harder again. Anyone can wax lyrical about extraordinary times, but running a wholesale or retail portfolio through them is a different job.
So I've settled on something a bit less clever. Build resilience to variance with the two tools every portfolio already has, liquidity and rebalancing. This is how I've been thinking about it, not a recommendation for anyone's portfolio:
- I'd hold more cash than the model says. Cash drag is the price of being able to act, as long as you don't chicken out and actually act!
- I'd get out of calendar rebalancing cycles and go on bands, because a quarterly review tends to arrive after the period that matters.
- Trend shouldn't be ignored. The aim is cash + 300-500bps, which looked pointless when cash was zero and looks sensible now that bonds have stopped hedging equities. It makes for a terrible client conversation though when trend has a bad month while markets rip, but the free lunch is in the aggregate.
- Credit funds with a negative one-year return look interesting, but the work is in balancing liquidity, quality, spreads and duration.
- You don't get paid to be a bear and the dislocations are where the returns are. I'd plan around them, with targets set, so that when markets sell off I can act.
Average
These are not average times, and the point of this piece is that I don't think you need to be certain about anything to act.
If I'm right, the average portfolio is geared to a regime that has ended, variance drag does the damage, and you find out slowly, one review at a time.
If I'm wrong, the cash costs a little and the trend allocation annoys a few clients, but the portfolio still captures most of the average.
Asymmetry is the whole point!