Where Should Your AI Agent Run? Cloud, or a Machine in Your Office

A small always-on cloud VM with 2 GB of RAM costs US$12 a month, roughly A$17 a month at A$1.44 to the US dollar on 29 July 2026, at both DigitalOcean and AWS Lightsail. The two providers have converged on the same price point. Cheaper tiers exist: DigitalOcean lists a 1 GB Basic Droplet at US$6 a month, and AWS Lightsail's smallest Linux bundle is US$5 a month for 0.5 GB. Those are the list prices published by each provider, checked on 30 July 2026. Treat the Australian dollar figures as arithmetic at one day's rate rather than a fixed price. Neither DigitalOcean nor AWS Lightsail publishes AUD list pricing for small VMs, so what you actually pay moves with the exchange rate and with whatever margin your card issuer adds on top. Over a year that drift is material. Check the rate on the day before you put a number in a budget. The office machine has a published Australian price with no conversion involved. A Mac mini with the M4 chip, 16 GB of unified memory and 256 GB of storage is A$1,299 from Apple Australia, and the M4 Pro starts at A$2,499. Treat A$1,299 as a floor rather than a realistic figure: 256 GB fills quickly once you are holding container images, model weights and run history, so a sensible agent box costs more than the base configuration.

An always-on Mac mini at idle costs roughly A$9.11 a year in inner Melbourne electricity. That is the calculation people skip, and it undercuts most of the argument they are having. Apple's Mac mini Product Environmental Report from October 2024 measures the M4 at 4.01 watts idle with the display on at 230 volts, 0.51 watts asleep and 0.11 watts off, with a power supply efficiency of 92.6 per cent. The price side comes from the Essential Services Commission's Victorian Default Offer for 2026-27, effective 1 July 2026. On the residential flat tariff, usage runs from 25.96 cents per kilowatt hour in the CitiPower zone covering inner Melbourne up to 31.98 cents in the AusNet zone in the outer east. A 4 watt device running continuously uses 35.1 kilowatt hours a year, which is A$9.11 at the CitiPower rate. One correction is worth making, because it is where most back-of-envelope sums go wrong. The daily supply charge, which ranges from 119.12 to 138.05 cents a day across the five Victorian distribution zones, is fixed. You pay it whether or not the server is plugged in, so it is not a marginal cost of running one and it must be excluded from any "what does this cost me" calculation. Only the cents per kilowatt hour usage rate is relevant. Heavier hardware costs more, and the following figures are illustrative arithmetic on assumed average draw rather than measured facts. Apple publishes only idle, sleep and off, and no sustained-load figure for the M4 Mac mini could be found. At an assumed 25 watt average, a busy agent with containers running, you get 219 kilowatt hours or about A$56.85 a year. At an assumed 60 watts, an older desktop under load, that is 525.6 kilowatt hours, or roughly A$136 to A$168 a year depending on your zone.

The crossover lands between about seven and nine years, which is longer than most businesses keep the hardware. The office machine costs A$1,299 once, plus somewhere between A$9 and A$57 a year in Melbourne power depending on how hard it works. The cloud VM costs about A$207 a year, being US$144 converted at A$1.44 to the US dollar. So the office machine saves you between roughly A$150 and A$198 a year, and A$1,299 of hardware takes seven to nine years to earn that back. That figure moves with the exchange rate. A weaker Australian dollar makes the cloud VM dearer and shortens the payback; a stronger one lengthens it. The arithmetic is not a fixed property of the two options, which is itself a reason not to let it decide anything. Seven to nine years is longer than most small businesses keep a computer, longer than most agents survive without a rebuild, and well past the warranty. That is the real conclusion from the cost arithmetic: the money is close enough to a wash that it should not decide the question. Something else should. One practitioner argued the opposite case on a Hacker News thread about overengineering a homelab, saying a continuous 60 watt draw would cost more than a full-time VPS. At the rate used here the arithmetic does not quite bear that out: 60 watts is A$136 to A$168 a year against about A$207 for the cloud VM, so the power still comes in under the cloud bill. It would flip if the Australian dollar weakened. Either way the two are the same order of magnitude, which is the point, and note that the commenter's figure is an estimate rather than research.

There are four genuinely different places an always-on agent can live, and they trade off differently on cost, control and who carries the risk. The table below is an analytical framing rather than sourced data, with the load-bearing prices sourced in the sections above. | | Cloud VM | Managed platform | Machine in the office | Proper on-premises | |---|---|---|---|---| | Up-front cost | near nil | near nil | A$1,299 and up | high | | Monthly cost | low and fixed, US$5 to US$12, about A$7 to A$17 | metered, scales with use | electricity only, A$9 to A$57 a year | electricity plus support | | Who patches the OS | you | the vendor | you | you | | Who carries power and internet risk | the provider | the vendor | you | you, with funded mitigation | | Data residency control | region choice | vendor dependent | total | total | | Run visibility built in | no, you build it | yes | no | no | | Realistic uptime | high | high | depends entirely on the office | high if funded | The row that decides most of these decisions is the fourth one. Cloud and managed hosting are, more than anything else, a way of buying somebody else's power and internet reliability. The row after it matters for a different reason: managed platforms are the only option where run history and error visibility arrive pre-built, which is why the office machine so often ends up invisible as well as fragile.

A consumer UPS buys you minutes, not hours, and riding out a two to three hour outage takes engineering no small business will maintain. The clearest account of this comes from a Hacker News thread on overengineering a homelab, where one commenter raised exactly the small business fear: their server could be shut down by a power outage while they were at work or on holidays, and a VPN could not solve that. The reply is the whole argument in miniature. Their mitigation escalated from a small UPS, to a UPS crudely modified with an angle grinder to accept a car battery, to a dedicated inverter, charger and transfer switch running off a caravan battery, to lithium power stations. They note the modified UPS eventually died. This is lived experience from one person rather than research, but it maps the shape of the problem accurately: each step buys more runtime and adds more to maintain. Another commenter on the same thread pushed back on the DIY route entirely, saying their commercial UPS already scared them for its fire potential and there was no way they would take on the risk of DIY work on something that could burn down the place or electrocute them. That is a reasonable position, and it is the honest answer to "can't I just put it on a UPS". You can. It will cover a flicker and a short interruption. It will not cover the outage that actually takes your agent down for an afternoon.

The failure mode of an office machine is not dramatic. It is that patching quietly stops happening and the system keeps appearing to work until the day it matters. Two commenters on a Hacker News self-hosting thread describe the arc directly. One said they had looped back around to not wanting to play sysadmin at home, and admitted most of the things they run had not been updated in a while. Another said the last thing they want to be doing on a Friday night is debugging why their home network is broken. Security follows from the same decay rather than from exotic attackers. On a thread about CLI agents making home servers easier, Simon Willison named security as the biggest reason he had not to run a home server, specifically that he was worried he might fall behind on updates and end up compromised. His mitigation is Tailscale, so no ports are exposed to the internet at all. The risk he is describing is his own discipline, not somebody else's skill. There is also a skills point that experienced infrastructure people make and software people tend to resist. One commenter, writing as an infrastructure engineer, put it plainly: the average software engineer can be great at writing code but not so good at managing a complex environment reliably. Those are separate jobs. This matters for a small business because the person who ends up owning the office machine is usually the web developer who built the site, or a family member who is good with computers. Neither has run production infrastructure, and neither signed up to.

Many Australian internet connections sit behind carrier-grade NAT with a dynamic IP address, which means inbound connections cannot reach a machine on that network at all. This is the most practically important objection to the office box and the least discussed one, because it is not a reliability problem you can improve with better hardware. It is a structural blocker. It matters enormously for agents specifically. Webhooks are how almost every agent gets triggered: a form is submitted, an order is paid, an email arrives, and the platform posts to a URL. If nothing on the public internet can open a connection to your office machine, that URL does not exist, and you need an intermediary such as a tunnel or a relay to make the agent reachable. That intermediary is now one more thing you run, and one more thing that can fail. Several commenters on the Hacker News home server thread raised this independently, noting that many ISPs use CGNAT and that a plain WireGuard setup cannot traverse it. One summarised the practical position well: they do not have a static IP, so Tailscale is convenient, and less likely to fail when they really need it than trying to deal with dynamic DNS. Before you buy hardware, find out whether your connection gives you a public IP address at all. Your provider can tell you in one phone call.

On the regulator's own data, Australian fixed broadband faults are common enough that an office connection is a weak link in an always-on system. The ACCC collects this under its Service Quality and Network Performance Record Keeping Rule. For the September quarter 2025, roughly 624,000 NBN service faults were rectified in the preceding twelve months, and emergency unplanned outages rose from 1,248 in the December quarter 2024 to 1,638 in the September quarter 2025. Around a million services were affected by Priority 1 faults in the March quarter 2025, which NBN Co attributed to seasonal weather. By access technology, HFC recorded the lowest average service level at 94 per cent across three consecutive quarters, with fixed wireless at 97 per cent. An important caveat applies to those two numbers: they were obtained via the ACCC's quarterly summaries rather than the underlying data, and how "service level" is defined was not confirmed. Do not read the 94 per cent as uptime or availability. Read it as a signal that HFC is the weakest performer on the regulator's own measure. The point is not that your connection is bad. It is that a cloud provider's redundant power and multiple upstream transit links are a different class of infrastructure from a single business NBN service and a power point, and you are choosing between them whether or not you think about it.

Most office machines do not fail for interesting reasons. They fail because somebody closes the laptop lid. The list below is stated from experience rather than sourced, because nobody publishes research on it, but every one of these has taken down a real system somewhere. The cleaner unplugs it to run the vacuum. The operating system installs an update overnight and sits waiting at a login prompt for a disk encryption password nobody in the office knows. The machine gets moved during an office fit-out and comes back on a different network. It gets stolen. The drive dies and there was no backup, because backups were on the list for after the busy period. The person who set it up leaves the business, taking the only working mental model of how it fits together. None of these are hard problems individually. Collectively they are the reason a cloud VM is the sensible default: not because the software is better, but because a machine in a data centre has none of these exposures. Nobody walks past it. Nobody needs the desk space.

There are five situations where a machine on your premises is the correct answer, and being clear about them matters more than winning the argument. First, a hard contractual or regulatory data residency requirement. Government contracts, some health data arrangements and some enterprise supplier agreements impose one. When it applies it is not negotiable, though note that Australian privacy law does not itself impose it. Second, genuinely high and sustained compute volume, especially if you are running an open-weight model locally rather than calling an API. A Mac mini's unified memory makes it a credible local inference box, and owning the silicon changes the arithmetic once volume is high enough. No break-even volume for local inference against API pricing could be sourced, so no threshold is stated here. If someone gives you one, ask where it came from. Third, avoiding metered pricing. Zapier bills tasks and Make bills operations. If the agent's work is high frequency and low value per run, such as polling, monitoring or bulk enrichment, the metered platform becomes the dominant cost and self-hosting is rational. Fourth, low latency to hardware physically in the building: a label printer, a till, a camera, a machine on the shop floor. Routing that through a cloud VM adds a round trip and a tunnel that can break. Fifth, data the owner does not want leaving the building even where the law permits it. Owner discomfort is a legitimate business input, not a misunderstanding to be corrected.

APP 8 does not prohibit storing personal information overseas. This is where most content on the subject is wrong, and the error costs owners money, because they buy hardware they did not need. The OAIC says so directly in Chapter 8 of its APP Guidelines on cross-border disclosure. Anyone telling a small business it must legally run its agent on a machine in the office is overstating the law. What APP 8.1 actually requires is that before disclosing personal information to an overseas recipient, you take such steps as are reasonable in the circumstances to ensure the recipient does not breach the Australian Privacy Principles. OAIC's guidance says this typically involves an enforceable contract, with provisions requiring APP compliance, binding subcontractors, providing a complaint process and covering data breach notification. The test is contextual: the sensitivity of the information, the nature of the relationship, and the consequences of mishandling. Section 16C is the sharp end. An entity that discloses personal information to an overseas recipient is accountable for that recipient's acts and practices as if it had done them itself, and that applies even where reasonable steps were taken. You can outsource the work. You cannot outsource the liability. There is one more distinction that does real work here. Routing or storing information with an overseas provider can constitute a "use" rather than a "disclosure", and APP 8 only bites on disclosure, provided the entity retains effective control of the information through binding contractual safeguards. So the honest line for a Melbourne business is: stay accountable for it and control it by contract, not keep it in Australia.

The Privacy Act generally does not apply to small businesses with annual turnover of A$3 million or less, but the list of exceptions catches more small operators than owners expect. OAIC's own page on rights and responsibilities states the threshold, and then lists the carve-outs. Private sector health service providers are covered regardless of turnover, and OAIC's list of what counts expressly includes complementary therapy and fitness services. A Melbourne personal training studio or a remedial massage practice turning over a few hundred thousand dollars is likely inside the Act, not outside it. Businesses that sell or purchase personal information are also covered, as are credit reporting bodies, contracted service providers under Australian Government contracts, and entities that have opted in. On reform, be careful about what you have read. Some commentary has run headlines saying the A$3 million exemption is gone. On the evidence available it is not: OAIC's page still states the A$3 million threshold and carries no repeal notice, and removal of the exemption sits in a proposed second tranche of Privacy Act reforms that has not been introduced as a Bill and has not been dated. What has already happened is narrower and targeted. Under the AML/CTF reforms, designated service providers including real estate, legal, accounting and conveyancing businesses have been within the Privacy Act for their AML/CTF data handling since 1 July 2026. If you are in one of those sectors, that change is real and it has commenced.

For almost every small business, put the agent on a cloud VM or a managed platform. The office machine is a false economy, and the price is not paid in dollars. It is paid the first time the power goes out during a long weekend, or the first time an operating system update leaves the machine sitting at a login prompt for three days while enquiries pile up unanswered. Here is the specific test. Would you notice within an hour if the agent stopped? If the answer is yes, because it answers customers or sends quotes or touches money, it should not depend on your office power point, your office internet connection and your own patching discipline. If the answer is no, the hosting decision genuinely does not matter much, and the attention belongs somewhere else. One more thing is worth naming, because it is the real motivation about half the time. If part of the appeal of the local machine was being able to see what the agent is doing, that is a different problem with a different solution, and buying hardware does not solve it. A machine under your desk gives you physical proximity, not visibility. Run history, exception alerts and an approval queue give you visibility, and they work identically whether the agent runs in Sydney, Singapore or the storeroom. That is covered in seeing what your AI agent is doing. If your concern is less about where the agent runs and more about who holds the records it creates, your records live on someone else's server is the companion piece.

Is it cheaper to run an AI agent on my own hardware? Only after seven to nine years. A Mac mini M4 with 16 GB is A$1,299 from Apple Australia, and Melbourne electricity for an always-on device runs between about A$9 and A$57 a year depending on load, using the 2026-27 Victorian Default Offer rates. A US$12 a month cloud VM is about A$207 a year at A$1.44 to the US dollar, so the office machine saves roughly A$150 to A$198 annually and takes seven to nine years to repay itself. That figure moves with the exchange rate. Can't I just put the office machine on a UPS? A consumer UPS covers a flicker and a brief interruption, not an afternoon. One Hacker News commenter described their own escalation from a small UPS, to a UPS modified with an angle grinder to take a car battery, to an inverter and transfer switch on a caravan battery, to lithium power stations, and the modified UPS eventually died. That is what riding out a multi-hour outage actually looks like, and it is not something a small business will maintain. Does Australian privacy law require me to keep client data in Australia? No. OAIC's APP Guidelines are explicit that APP 8 does not prohibit storing personal information overseas. The obligation is to take steps reasonable in the circumstances to ensure an overseas recipient does not breach the APPs, which OAIC says typically means an enforceable contract. Under section 16C you remain accountable for that recipient's conduct as if it were your own, even where you took reasonable steps. Why can't webhooks reach a machine in my office? Many Australian connections use carrier-grade NAT and a dynamic IP address, so nothing on the public internet can open an inbound connection to your machine. Webhooks are how almost every agent gets triggered, so this is a structural blocker rather than a performance issue. You would need a tunnel or relay as an intermediary, which is another component to run and another thing that fails. Has the small business privacy exemption been removed? Not on the evidence available. OAIC's page still states the A$3 million turnover threshold with no repeal notice, and removal sits in a proposed second tranche of reforms that has not been introduced as a Bill and has not been dated. What has commenced is narrower: since 1 July 2026, AML/CTF reforms have brought designated service providers including real estate, legal, accounting and conveyancing within the Act for AML/CTF data handling. What is the strongest reason to run an agent locally? A hard contractual or regulatory data residency requirement, because it is the only one that is not negotiable. After that, sustained high compute volume where you are running a local model rather than calling an API, avoiding per-task or per-operation metered pricing at high frequency, and low latency to hardware in the building such as a label printer or a till.