SharonAI

IThe business

SharonAI Holdings rents out computing power built on NVIDIA chips, mostly to companies that train and run AI models. It calls itself an Australian neocloud.³ In plain words, it purchases graphics chips, installs them in other people’s data centers and sells time on them, usually under contracts of five years or more. It does not own the buildings or the power plants. The head office is in New York; the staff and the equipment are in Australia, and from next year in New Zealand.³¹⁶ The year is the calendar year.

The company is young and its history is short. SharonAI Inc. was formed in Delaware in February 2024. It folded in an Australian company that owned old crypto-mining servers, bought a data-storage business for the Filecoin network in June 2024, and walked away from Filecoin a year later to rent GPUs instead.³ In January 2025 it set up a joint venture with New Era Energy and Digital to build a 250MW data center in Texas, and sold its half back in January 2026 for about $74M in cash, shares and a note.³⁵ It became public in December 2025 by merging into a listed shell, Roth CH Acquisition Co., and did a 1-for-50 reverse split on 23 December.³ It listed on Nasdaq in February 2026 and raised $125M before costs.³⁵ So there is no ten-year record here. There are two audited years, 2024 and 2025, and two quarters of 2026.

What it sells today is small. Revenue in the second quarter of 2026 was $1.9M, from renting GPU time, against $0.4M a year earlier.³ Over the twelve months to June it was $3.1M C (calculated).²³ The first customers named in the filings are Canva and GMI Cloud, signed in February 2026.⁵ The machines in service are a cluster of more than 1,000 chips at a NEXTDC data center in Melbourne.⁵ The company also says it sells in two ways, on demand and under take-or-pay contracts, where the customer pays for a fixed amount of capacity whether or not it uses it.² The size of the business is in the contracts, not in the revenue.

The contract book

Between April and August 2026 SharonAI announced five contracts. The company’s own total, which it calls total contract value, is about $8.8B.⁴ The table lists them with what each needs in hardware, from the 10-Q and the S-1.

The five do not sit alike. Four are ordinary leases of chips to a customer for five years. The fifth, with NVIDIA, is four times larger than any other and is described differently. The 8-K says the agreement is worth “up to” $4.88B, that all the phases of the build must end on the same date whenever each is delivered, and that some capacity reserved for NVIDIA may be sold on to third parties, with a share of that revenue going to NVIDIA.⁸ The press release describes a revenue-sharing and credit-support structure, in which NVIDIA earns ordinary product revenue as the chip supplier and also a share of the cloud revenue.¹¹ I come back to what that means for the numbers in Section VI.

Where the chips go

The company says it has secured 212MW of data center capacity, of which 120MW is contracted, and that more than 68,000 NVIDIA chips will be running by mid-2027.¹⁰¹² NEXTDC, an Australian data center operator, hosts substantially all of the equipment, and in July the company signed an amended order for 72MW with it, a commitment of about $623M over the term, in five phases from April to August 2027.³⁵ The 10-Q does not give that term. A further 15MW is contracted from GreenSquareDC in Australia, and 14MW in New Zealand for the AI lab.³ Other pieces come from named suppliers: servers from Lenovo and World Wide Technology, storage software from VAST Data, and an orchestration platform from Rafay Systems under a five-year agreement signed in September.³¹⁷

The first of the large deployments has been delivered. On 20 August the company said the global technology customer had accepted the first phase, which released cash held in escrow.¹⁰ It is the only delivery milestone of the new contracts on the public record so far. The revenue from the others starts next year.

Who pays for it

Customers pay the company, and the company has to pay for the hardware first. That is the business model and its main risk. In the first half of 2026 SharonAI spent $323M on equipment and deposits against $2.2M of revenue.³ It paid for that with money raised from investors: $125M at its listing, $350M of 6.00% convertible notes in May, a $900M sale of shares and pre-funded warrants in June at $68.73 and $700M of 4.75% convertible notes the same month, and on 1 October a $356M loan secured on the chips at a fixed 9.95% for most of it.⁵³¹⁸¹² At 30 June it held $1.86B in cash.³ It had commitments of about $6B and signed about $1B more in July.³ The shortfall is large, and I size it in Section VI.

IIThe moat

I ask two questions. Could a rival take the customer at the same price, and would the customer move? For a business that rents chips, both answers lean the wrong way for the owner, and the filing says so more directly than I expected.

The product is a commodity

A GPU hour is a GPU hour. Customers care about price, availability, how soon the chips arrive, and whether the data stays in the country. The 10-K says many competitors are larger operators with better access to power and equipment, that SharonAI competes with them for equipment, energy and capital, and that the technology can be made obsolete by newer chips; the 10-Q adds that GPU supply is constrained and costly.²³ NVIDIA has been bringing out a new generation of chips roughly every year, so a contract signed for 2027 prices a chip that will be several generations old by 2031 E (estimate). Price per chip-hour for older chips falls; that is the main reason I assume renewals come at less than half the original price in my base case E (estimate).

What it does have

Three things stand between SharonAI and a plain commodity. First, power and floor space. Capacity in good data centers is hard to get, and the company has contracted 212MW, much of it through NEXTDC.¹⁰⁵ Second, supply. It is a certified NVIDIA Cloud Partner, NVIDIA supplies its chips, and under the June agreement NVIDIA shares in the revenue and gives credit support.¹¹⁸ Chips are scarce, and a supplier that wants you to succeed is worth something. Third, sovereignty. Government and research customers in Australia and New Zealand may prefer to keep sensitive data and workloads inside the country.³ That preference is real, though I could find no filing that shows customers paying a premium for it.

Where it is weak

Almost all of the equipment sits in one landlord’s buildings. The S-1 says substantially all of the GPU infrastructure is in NEXTDC facilities and that a problem at NEXTDC could hit all or most of its capacity at once.⁵ It also says its customer contracts do not always mirror supplier terms. Lenovo may end service if there is a material adverse change in SharonAI, and the customer contracts do not clearly give SharonAI the same right against its customers.⁵

The customers are few and large. Three of the five contracts are with firms the company has not named, and the S-1 says the business has and is expected to keep significant customer concentration.⁵ A customer that is most of your revenue has the power to reprice the next contract. And the competition is real: the 10-K says many rivals are larger and better placed on power and equipment.² I have not tried to measure how they price in Australia; my point is only that this is a field where anyone with a good contract can raise money.

My judgement: SharonAI has a good position in a market that is short of supply, and a few real advantages, scarce power and the supplier link. It does not have a wall around the customer. The return it earns will be set by what it paid for chips and how long they keep earning, not by loyalty.

IIIThe competitive advantage

The advantage here is speed with money. In ten months SharonAI went from a company with a few million dollars of revenue to one that has raised over $2.6B and signed about $8.8B of contracts.¹²⁴ Few young companies can do that, and in Australia few rivals have tried. The relevant question is whether being quick to sign is the same as being good at the economics of what is signed. The contract table says it may not be.

How the contracts pay

Take each contract and ask how long the customer’s payments take to repay the hardware. I use the contract value, spread evenly over the term, and take off 22% of revenue for data center space, power, staff and software, which I judge to be a fair central cost E (estimate). If NEXTDC’s 72MW order serves the NVIDIA fleet, its $623M is 13% of the NVIDIA value, and there are other costs on top.³ The result is how many years of the term it takes to earn the hardware back.

Read that for what it is. Four contracts earn their hardware back with one to two years of a five-year term to spare, which is a return of roughly 11% to 18% a year before tax on the cash put in, fine if the customer pays and the chips last, and no more C (calculated). The fifth, which is two-thirds of the capital, does not earn its hardware back inside its term at the disclosed value. Either the “up to” value understates what the fleet will earn, because NVIDIA’s reserved capacity can be sold on to others, or the programme is a poor use of $4B. The filings do not say which. The press release says the structure lets SharonAI scale with less capital of its own.¹¹ That could mean NVIDIA’s support reduces what SharonAI has to put in. It does not change what the 10-Q says is committed.

What looks good

The best contract on paper is also the one that has been delivered. The global technology customer pays $950M for hardware that costs $468M, 2.0 times, and has accepted its first phase.⁵¹⁰ Lenders have also committed money against the book. The $356M facility, which Goldman Sachs and private credit funds joined, is secured on the chips and on one customer contract, and its first $150M tranche is to be repaid from a customer deposit held in escrow once the customer accepts delivery.⁶¹² The 8-K does not say which contract. A lender that takes security over the chips and the contract is lending against that contract’s cash flow, and that is a vote of confidence in one contract, not in the whole book.

The second thing is the funding record. Since December 2025 the company has raised money from convertible-note investors, from Oaktree and other funds in the notes and the June placement, and now from banks and credit funds.¹⁹¹⁸¹² Raising money is part of this business, and a company that can do it at the right price has an advantage over one that cannot. Whether the prices were right is a different question, and I take it up in Management.

Against the alternatives

The other neoclouds are the nearest comparison; CoreWeave and Nebius are on this site. Their model is the same: sign long contracts with a few large customers, borrow against them, and purchase chips. SharonAI’s differences are size, which is far smaller, and the fact that all of its revenue is still ahead. A reader who knows the larger names will recognise the risks listed in Section VIII: the funding gap, the customers, and the useful life of the chips. I have not tried to rank them.

IVManagement

James Manning, 42, is the chief executive and a co-founder. He became chief executive on 22 January 2026; before that Wolfgang Schubert held the post, and Manning was chairman.⁵⁷ The chairman is now Andrew Penn, a former chief executive of Telstra, who joined the board in May.⁵ Manning founded and ran Mawson Infrastructure Group, a listed digital infrastructure company, until May 2023.⁵ Two other senior people, the former finance chief Tim Broadfoot and Nicholas Hughes-Jones, who runs corporate development, also worked at Mawson.⁵ Andrew Leece, the third co-founder, came from Macquarie Bank and aviation finance.⁵ Between July and September the finance chief, the operating chief and the legal chief were replaced or newly named: Anuj Goel took over finance on 24 August, David Burns became operating chief on 7 September, and Melissa Anastasiou, from Spark New Zealand, became chief legal officer on 23 July.¹³¹⁴⁵ Leece moved to head of strategic partnerships.¹⁴ On 20 September the company told HoganTaylor, its auditor for 2024 and 2025, that it would not continue, and on 23 September engaged Ernst & Young.¹⁵ The filing reports no disagreement and no qualified opinion.¹⁵ A company this size getting a bigger auditor and a bigger team is normal. Changing both in one quarter, as the biggest contracts begin to be delivered, is a lot to ask of a young organisation.

How they are paid

Until this year the pay was small. In 2025 the former chief executive earned $302,000, the finance chief $221,000 and the operating chief $213,000, and Hughes-Jones $1.18M, of which $1.13M was listed as other compensation.⁷ The new packages are larger and mostly in stock. From 1 May 2026 Manning is paid a base of A$704,225, about US$500,000, with a short-term incentive of up to 200% of base, a long-term incentive of up to 150% in restricted stock units, and a one-time listing award of 25% of base.⁷ The new finance chief gets A$650,000 with up to 100% short-term and 200% long-term incentives; the operating chief A$550,000 with up to 150% and 150%.¹³¹⁴ The consulting contracts through which the founders were paid separately were ended in April 2026.³

What does this pay them to do? The proxy says each restricted stock unit carries a key business performance target and vests 12 months after the target is met.⁷ I could not find the targets in any filing. Without them I cannot say whether the pay rewards cash returns or only growth in contracts and capacity, which the company reports as total contract value and megawatts. Both are easy to grow by signing deals that cost more than they earn. I expect the next proxy statement to list the measures.

What they own, and how they vote

The founders own a lot of stock and carry more of the votes. At the proxy record date Manning held 1,562,984 shares, Leece 1,446,175 and Hughes-Jones 1,285,564, about 4.4%, 4.1% and 3.6% of the Class A stock; together with other officers and directors the group held 12.99%.⁷ At $22.68 the three founders’ shares are worth about $97M C (calculated). But there is also a Class B stock, 136,341 shares split equally among the three founders, with 160 votes each. That gives the officers and directors as a group 45.94% of the votes on 12.99% of the shares.⁷ The founders can carry any vote that most of the other holders would not unite against. The record date was 2 July, before the last pre-funded warrants were exercised on 15 September, so the share percentages are now a little lower; the vote share is also a little lower, because the Class B count did not change.

They agreed to lock-ups on their shares to 31 March 2027 when the 6.00% notes were sold.¹⁹ There are six Forms 4 on file for 2026, four from officers and directors and two from Situational Awareness, and none reports an open-market sale.²⁰ One records 90,893 shares issued to a Hughes-Jones family trust to settle an old reimbursement obligation.²⁰³

One more fact on the owner side, stated as a dated fact. At 30 June 2026 Situational Awareness LP held 5,396,127 shares, worth $456.8M that day and 2.26% of what it reported in total, according to its 13F-HR filed on 14 August.²¹ The company’s press release of 17 June named the fund as an anchor of the June placement.¹⁸ A 13D filed on 28 August and amended on 17 September shows 8,070,950 shares, 21.1% of the Class A stock, after the fund exercised its last 2,674,823 pre-funded warrants on 15 September.²²²³ These facts show that a large holder exists and where it stood on those dates. They are not a reason to value the stock, and I draw no conclusion from them.

What they did with capital

There is no full cycle to judge, only about two years of history, so I look at what was raised and at what price. The earliest outside money came in December 2025, as $103.7M of convertible notes.³ They converted on 11 June 2026, under a formula tied to the listing, at an effective $12.53 a share, into 8.43M shares, when the stock was about $71.51.³ The shares issued were worth about $603M on the conversion date, so holders turned $104M into roughly 5.7 times that in six months C (calculated). The company recorded a $471M non-cash loss for the rise.³ Insiders took part on the same terms as outsiders: companies tied to Manning, Hughes-Jones and Leece subscribed for about US$1.46M of the notes.³ That is a generous deal for the first money in, and it cost the other holders 8.4M shares, about 22% of the 38.5M outstanding today C (calculated).

After that the selling was well timed. The company raised $350M of 6.00% notes convertible at $48.24 in May, then $900M of shares at $68.73 and $700M of 4.75% notes convertible at $99.66 in June, the day before its highest close.³¹⁸ The stock has since fallen to $22.68, 67% below the June placement price. Selling shares near the top is a good decision for the holders who stay, and management made it. The 4.75% notes, convertible at 45% above the placement price, are low-cost money if they are never converted: $700M at 4.75% for six years.¹⁸ By October, money secured on the chips cost 9.95%.⁶

The record is weaker on commitments. By 30 June the company had committed about $6B to equipment and data center space, before it had delivered a large contract or generated more than $2M of revenue in a quarter.³ On 21 August it obtained the consent of holders of the 6.00% notes to remove the limits on its borrowing and on liens, so more secured debt can now rank ahead of those notes.²⁴ This is the kind of change a company makes when it expects to borrow a lot.

Following the incentives, then: the founders are paid in stock and own a lot of it, and the Class B shares let them keep control. That pays them to grow the company. It does not pay them to be careful about which growth is worth the capital, and the record so far is one of fast signing and fast fundraising, with the returns on the equipment still to be shown. I give the funding speed credit, and I note that the best-timed decisions were the share sales.

VFinancials

The record is too short for ten years and too young to read as a trend. SharonAI Inc. did not exist until February 2024. The audited statements cover 2024 and 2025, and the 10-Q covers the first half of 2026. They are the statements of a company that is changing what it does every few months: it ran a data-storage business for the Filecoin network, mined small amounts of digital assets, built a stake in a Texas data center and sold it, and only in 2026 began renting GPU time at a meaningful scale.³ I show what exists and no more.

Revenue has risen from $0.4M in 2024 to $1.6M in 2025 and $3.1M over the twelve months to June C (calculated).²³ The gross margin is not informative yet: the second quarter had $1.2M of gross profit on $1.9M of revenue, and the quarter before it had a gross loss.²⁵³ The company’s adjusted EBITDA for the quarter was $0.6M, against a GAAP operating loss of $24.8M. It gets there by adding back $14.3M of unrealized foreign exchange loss, $2.7M of stock pay and $1.9M of depreciation, among other items.⁴

What an owner could really take out

Nothing. The headline net loss for the first half of 2026 is $450M, and $471M of it is one accounting entry: the change in the fair value of the December 2025 convertible notes between the start of the year and their conversion in June.³ It is a loss because the stock went up. There is also a $66M gain from selling the Texas joint venture.³ Neither is a measure of what the business earns, which is close to nothing so far.

The cash picture is different, and larger. Operating cash flow in the first half was a positive $23M, which looks good until you see that $144M of it is customer deposits, cash received for service not yet given.³ Without it operating cash flow is −$121M C (calculated). Add $323M of equipment spending and prepayments and $3M of stock pay, and owner cash flow for the half was −$303M as reported, or −$447M without the deposits C (calculated). For the twelve months it is −$316M and −$460M. The deposits are real money, and they have to be earned back by delivering service, so I set them aside.

How much of the spending is needed just to stand still? Nearly none yet. The business is not standing still. It is being built, and that is why owner cash is so far below any earnings measure. What matters is what the equipment will earn once built, and the next section turns to that.

Quality of earnings

Three items need adjusting before the profit figures mean anything. The first is the convertible-note entry above. The second is the $66M gain on the Texas sale. Both sit below the operating line, so the operating loss of $27.6M for the half is cleaner. The third is inside it: “other expenses” of $14.6M in the second quarter, which the company says came mainly from unrealized foreign exchange losses on remeasuring foreign-currency balances.³ The cash flow statement puts the half-year figure at $11.1M.³ Cash held in one currency by subsidiaries that report in another creates those, and they move with the exchange rate. Take them out and the operating loss for the half is about $16.5M, almost all overhead for a company that had $2M of revenue C (calculated).

Three small errors are worth a line, because they say something about the finance function. The 10-Q gives the price of the pre-funded warrants as $68.2799 and the press release $68.7299, which is the placement price less the $0.0001 exercise price; the $900M total supports the press release.³¹⁸ The results press release shows a second-quarter loss per share of $26.16 where the 10-Q shows $22.84.⁴³ And a 10-Q/A of 11 August corrected 2025 per-share figures.²⁶ None affects my numbers. All three are the kind of slip a young finance team makes, and the finance chief changed in August.

The balance sheet

At 30 June SharonAI held $1.86B in cash and $13M in certificates of deposit against $1,007M of convertible notes on the books, which have a face value of $1,050M, $350M due in 2031 and $700M in 2032.³ Customer deposits of $144M are also owed in service.³ After those, net cash is $675M, or $17.13 a share on 39.4M shares C (calculated). But the cash is spoken for. The company has paid $303M in prepayments and committed another $5.9B, and it signed about $1B more after June.³ On 1 October it arranged a $356M loan against the chips at fixed rates of 7.25% rising to 9.95% for Facility A and 9.95% for Facility B, due in a lump sum 42 months after first drawing, with a covenant on how much it can borrow against the equipment’s value.⁶ The facility becomes non-recourse to the parent once the smaller tranche is repaid, apart from a management agreement.⁶ I read that as sensible structuring, and also as a sign that the next billions will be borrowed at about 10%.

The gap, as I count it, is $3.6B: commitments still to pay of $5.9B, less $1.87B of cash, less the $356M loan C (calculated). Customer payments from 2027 will pay some of it down, and the company says it expects the loan to be the first in a series.¹² The 70% to 80% borrowing against equipment that the S-1 targets would cover most of the hardware, leaving an equity piece the company has not yet said how it will raise.⁵ If it were raised as shares at $22.68, it would be much more dilution than the same raise would have been in June.

VIValuation

This is hard to value, and I would rather say how hard before I give a number. The company has revenue of $3.1M, contracts of $8.8B, and hardware commitments of $6.2B. The value is the small, uncertain difference between those two large numbers, after tax, overhead and the cost of the money. Four points on running costs, or three on the discount rate, move my base by more than its whole value. I give a base, a bear and a bull, and I show how far each depends on the few things I cannot see.

What would most likely destroy this

Before the numbers, the ways it goes badly wrong. I see three.

The money runs out before the cash arrives. SharonAI has paid $303M of prepayments and owes about $5.9B more for hardware, against $1.87B of cash and deposits and a $356M loan, so about $3.6B is still to be found, before 2027 receipts and before $1,050M of convertible notes fall due in 2031 and 2032.³⁶ In my base case, with interest at 11% on what is borrowed, the borrowing peaks at about $4.8B in late 2027 E (estimate). The shares have fallen 76% from their June high and secured money now costs 9.95%.¹⁶ If lenders and investors in new shares stop, the company cannot take delivery of what it has ordered, and the customers can walk. The gap is being closed as I write, and I do not know how much has been closed since June.

The contracts are worth less than their headline. Three of five customers are unnamed. ESDS is an Indian cloud company with ₹472 crore of revenue in its last financial year, which at about ₹85 to ₹90 to the dollar is some $55M, against payments to SharonAI of about $250M a year; it opened a ₹720 crore IPO on 28 August, with ₹576 crore earmarked for its own cloud equipment.²⁷ I do not know how it will pay. The S-1 says ESDS must post $140M in letters of credit or bank guarantees, and pays monthly in advance in US dollars.⁵ A short seller’s summary says that is larger than ESDS’s total assets of about $69.5M; I could not verify that figure.²⁸ The NVIDIA agreement is “up to” $4.88B, can be ended for frequent chip unavailability or an adverse change in SharonAI’s finances, and shares revenue with NVIDIA.⁸ A news site, citing Reuters, reported on 28 August that NVIDIA had paused parts of a financing initiative that exchanged credit support for a share of cloud revenue; the article carries no SharonAI statement, and I could not open the Reuters original.²⁹

The hardware earns less than it costs. At the disclosed values the NVIDIA programme earns back 0.95 times its hardware inside its term, before tax C (calculated). That is the whole case. If it is right, the biggest contract destroys value; if it is wrong, because spare capacity is sold on at good prices, it is the main source of value. The press release says NVIDIA earns a share of the cloud revenue, and the 8-K says some reserved capacity may be sold to third parties. Neither says how much, or at what price.¹¹⁸

What an owner could really take out

Nothing so far, as the Financials section shows: owner cash flow was −$316M over the twelve months to June as reported and −$460M without customer deposits, against a reported net loss of $486M that is mostly an accounting entry C (calculated). In my base case it turns positive in 2028 once the hardware is in service, runs at about $1.3B a year before financing, and is almost all spent repaying the borrowing until about 2033. After that the contracts end, the chips are four to seven years old, and what remains is renewals at lower prices. The cash an owner can take out is small, and late.

What I cannot judge

I cannot see the revenue per chip on the NVIDIA programme, or how much of the spare capacity will be sold. I cannot see the cost of power and space beyond one number, NEXTDC’s $623M for 72MW, and I do not know its term.³ I cannot see what the company holds in cash today: the last statement is 30 June, and third-quarter results are due by mid-November.³ I cannot judge what a four-year-old GB300 will rent for. And I cannot tell whether ESDS or the three unnamed customers will pay. For these I show ranges, and I do not pretend the base is more than a central case among wide ones.

How I value it

I build each contract by quarter from its stated value, term, hardware cost and start date, take off running costs, overhead and 30% tax with losses carried forward, and discount the owner cash flow before financing at 11% in my base. The cash and notes at 30 June are then added and subtracted, and the answer is divided by the 39.4M shares now outstanding including employee awards. Money borrowed or raised to fund the gap is assumed to cost the discount rate, so it neither adds nor subtracts value; if new shares are sold above my value, existing holders gain, and below it, they lose. The 11% is a blend. The Treasury 10-year yields 4.94%.³⁰ Lenders charge 9.95% against the chips, and the company’s equity would cost much more E (estimate). I model no new contracts in the base, because a contract that earns exactly its cost of capital adds no value. The convertible notes are treated as debt unless the value per share is above their conversion price, $48.24 and $99.66, in which case they become shares.³ I count the notes at their $1,050M face value. At 11% their coupons and principal are worth about $825M today, so face value is conservative by roughly $5.70 a share C (calculated). For the NVIDIA fleet, running costs are charged on the revenue actually received, so unsold chips cost nothing in my model; the fixed side is NEXTDC’s $623M for 72MW, about $104M a year if it runs six years, against the $228M a year I charge in the base. For the other four contracts the costs are charged before any haircut on revenue. The cases:

Bear

The NVIDIA fleet is delivered two quarters late and pays 85% of its stated value, after availability penalties and the revenue share. None of its spare capacity is sold on. ESDS pays for about half of its term. The company has to fund the hardware in the market at 14%, and when the contracts end the chips are worth little. The hardware costs more than the contracts bring in, and nothing is left for the shares: the model says less than zero, and a share cannot be worth less.

Base

The five contracts are delivered on time. ESDS pays 85% of what it owes, a risk allowance and not a forecast, and the other three pay 95%. NVIDIA pays its stated value, and the company sells about 6,200 more chips’ worth of capacity, 35% of what the 40,000-chip fleet could earn above the contract, at the price the AI-platform customer pays. Running costs are 22% of revenue. The chips renew for two more years at 45% of the original price. Money for the gap costs 11%. Owner cash turns positive in 2028 before financing, and the borrowing is repaid by about 2033.

Bull

The NVIDIA fleet fills: 70% of the spare capacity is sold, about 12,400 chips’ worth, and NVIDIA’s share of that does not eat the margin. All the other contracts pay in full. Running costs fall to 19%. Chips renew for three years at 65%. The company wins three more tranches of $1.5B hardware in 2028, 2029 and 2030 on terms like the best contract today, and the 6.00% notes convert into shares because the value is above $48.24. The rate is 9.5%.

What the price says

Turn it round. At $22.68 the market values the shares at $894M. Add the notes, deposits and leases and take off June’s cash, and the market is paying about $219M for five contracts worth $8.8B on paper and for everything the company hopes to build. My base value of the same, after the hardware, tax and overhead, is −$486M E (estimate). So the price needs about $705M more than my base finds.

The most useful way to read this is the NVIDIA fleet. For the price to be justified, with everything else at my base, the company has to place 83% of the capacity above the contract. The fleet then earns about $1,347M a year, or $3.84 for each chip an hour, close to the $4.15 the AI-platform customer pays for an older B300 chip.⁹ Every one of the 40,000 chips would have to be working, none of the revenue would go to NVIDIA, and the whole thing would have to start on time. That is possible. It is not a base case.

How much the answers move

The surprise is how large the swings are against how small the base is. Moving the share of NVIDIA spare capacity from 35% to 50% adds about $6 a share, which is more than the base value itself C (calculated). At no extra sales the base falls to −$9 a share, shown as zero. A business that earns a thin margin on a large pile of capital is that sensitive, and one more reason I show a range from $0 to $63 and do not claim to know where in it the shares belong.

The margin of safety

I cannot set it from the bear case, because the bear case is zero. No price is safe against it. So I put the cushion where the base is least certain, at half the base: $2.40 E (estimate). At that price the shares would be worth about $95M in all, a seventh of the $675M of net cash after the notes, though that cash is already committed to hardware. The shares have not traded that low since listing: the lowest close was $17.52 on 3 March.¹

The price against the range

The price is 373% above my base and 64% below my bull. At $22.68 my base cash flows return −8.7% a year to October 2036, my bull 27.3% and my bear loses everything E (estimate). Weighting the cases 25%, 50% and 25% gives $18 E (estimate), below the price. That weighting is generous: a bull this far above the base deserves less than a quarter of the weight, because it needs the NVIDIA fleet, the renewals, the follow-on contracts and a low cost of money all at once. The market may be right and I may be wrong. The five contracts can still turn out to be as good as the best, and the cash is real: $1.87B at 30 June against a market value of $894M. I would rather show how much has to go right than move the numbers until they meet the price.

Against the alternative

Every dollar here could go elsewhere. A ten-year Treasury pays 4.94%.³⁰ My base cash flows return −8.7% a year here. The nearer comparison is on this site. My Oracle page, another company spending heavily on AI computing, finds its price 7% above my base and a base return of about 11.4% a year. Oracle has revenue, customers, and a record, and SharonAI has contracts. I make no comparison with the other GPU clouds here; the CoreWeave and Nebius pages use their own prices and rates.

Quality and price, separately

How good is the business? I cannot yet tell. It has signed large contracts quickly and financed them, and its returns on the hardware, on the figures it has disclosed, are modest to poor. How good is the price? On my numbers it is high, with the cash a real cushion but already committed. A fair summary is that the price is for a business whose quality is still unproven, and that the shares have fallen by three quarters from a price that assumed more.

How the odds multiply

The price needs five things to go right together. My guesses are in the table. They are not independent, since a fleet that is delivered on time is also likelier to be sold on and renewed, so the true chance of all five is higher than the product. But each step is a coin toss or worse, and the product is about 3%.

Where this could be in 5 and 10 years

The year-10 value is zero in every case, and that is a limit of my model, not a finding about the company. The model holds only the five signed contracts and their renewals. By October 2036 almost all of that revenue has run off, and what is left does not cover overhead. Whatever was worth taking out has already been paid out, which is the cash column: $11.1 a share in the base. A real company would keep buying chips and signing contracts, and what it would be worth in 2036 depends on what that new money earns, which I cannot estimate. So read the year-10 column as zero by construction, read the return a year as the return from these contracts alone with nothing counted after 2036, and take the year-5 value and the cash paid out as the numbers with content.

The base pays out $11.1 a share by 2036, about half of today’s price, so it returns −8.7% a year. It needs the NVIDIA fleet to place a third of its spare capacity, money at about 11% and renewals at 45% of the original price. The bull pays $117 because the fleet fills, costs fall and the company adds $4.5B of good contracts. The bear never repays what it borrows.

The most likely way the base is wrong is the NVIDIA programme. Its stated value does not repay its hardware, and I have credited it with sales that no filing proves. With none of them the base is below zero.

These are estimates of value, not forecasts of the share price, and the market price can sit above or below value for years.

VIIThe case for the price

What most people miss

Three things cut in SharonAI’s favour and rarely come up. First, the shares are worth less than the cash. At 30 June the company held $1.87B in cash and deposits, against a market value today of $894M C (calculated).³¹ After the notes and the customer deposits, net cash is $675M, $17 a share, about three quarters of the price C (calculated). It is committed to hardware, but the hardware is what the contracts pay for. Second, the best contract is a good one: $950M for $468M of hardware, with its first phase accepted, and lenders have committed $356M against a contract in this book.⁵¹⁰⁶ Third, the NVIDIA 8-K says reserved capacity may be sold to others, the one place the headline value may understate what the chips earn.⁸

The bull case

Demand for AI computing is running ahead of supply, few providers serve Australia and New Zealand, and a young company with scarce power capacity and a supply tie to NVIDIA has signed five contracts in five months. Management says it is on track to bring contracted capacity online as scheduled and expects revenue to ramp materially from the third quarter of 2026 through 2027.⁴ Oaktree led the $350M of 6.00% notes and was named an anchor of the June placement, according to the company.¹⁹¹⁸ B. Riley started coverage on 9 July, and Northland Securities has a positive rating and a $95 target from 18 September.³¹³² On 30 October six analysts’ targets averaged $104.50, from $90 to $124, four to five times the price; I could not open their notes or match targets to firms.³¹

Not everyone agrees. Bleecker Street Research published a short report on 30 April, questioning the chief executive’s past company and the quality of the contracts, and says it is short the companies it covers.³³ I could not open it. A summary says the main attack was on ESDS, and reports no reply.²⁸ Since then SharonAI has signed four more contracts and had one accepted, and I have found no filing that answers the point about ESDS’s ability to pay.

Against reality

Today SharonAI bills $1.9M a quarter. If it delivers the 64,000-plus chips from April 2027, collects from ESDS and the unnamed customers, funds the gap near 10%, and places 83% of the NVIDIA fleet’s spare capacity at the AI-platform price, the price is possible: the fleet would then earn $3.84 a chip-hour on average, against $2.32 in the contract C (calculated). My base needs about $390M of revenue a quarter by the end of 2027, 200 times today’s, and the evidence so far is one accepted phase of one contract.

What a 10x would take. Ten times $22.68 is $226.80. With both note issues converted, the count would be 53.7M shares and the market value $12.2B, 14 times today’s, more than ten because the notes become shares C (calculated). That is 3.6 times my bull value. At a generous 10 times owner cash, it needs about $1.2B a year. A fleet like this keeps only about 14% of revenue as owner cash once the chips have to be replaced: a 78% margin, less replacing the hardware over five years, less tax E (estimate). That is about $8.7B of revenue a year, 5.0 times my base peak and 2.3 times my bull peak. Tenfold in five years is 58% a year, in ten years 26%. It has happened to a few companies, not to one that bills $3M a year on thin hardware margins. From here it is close to impossible on my numbers, and I would not build a case on it.

Where I come down

What moved me: the speed of the financing, a first accepted delivery, and a contract at 2.0 times its hardware cost. What didn’t: NVIDIA economics that do not repay the hardware at the disclosed value, an ESDS contract that is large against what I can see of ESDS, and a structure that lets three founders carry the vote. What would raise my values: evidence that the NVIDIA fleet sells at B300-class prices, and the next $3B funded at about 10% with little new equity. Both are likelier to show in the next two quarters than the next two years.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• NVIDIA is supplier, customer and partner. It supplies the chips, owns the largest contract, shares in the revenue and gives credit support.⁸¹¹ A news report of 28 August, citing Reuters, said it had paused some similar financing.²⁹ A change in how NVIDIA treats smaller cloud partners would reach SharonAI on all three sides.
• The chips age faster than the contracts. A contract term of five or six years is long for a chip that is replaced about every year. I assume renewals at less than half the original price, and a customer is free to ask for new chips instead.³
• One landlord. Substantially all the equipment is in NEXTDC buildings, and the S-1 says a problem at NEXTDC could hit most capacity at once.⁵ The NVIDIA phases depend on NEXTDC delivering space from April 2027.³
• Delivery. The company has delivered one phase. Over the next nine months it has to take delivery of and commission more than 60,000 chips across Australia and New Zealand, with a new operating chief and finance chief.⁴¹⁴¹³ Its own risk factors say a delay in hardware or space could leave it liable to customers.⁵
• Control and conduct. Three founders carry about 46% of the votes on 13% of the shares.⁷ The consulting arrangements that paid them separately ended in April.³ The auditor changed in September, the press release and 10-Q disagree on the quarter’s loss per share, and a short seller has raised questions I have not been able to test.¹⁵⁴³³
• Dilution. The 6.00% notes convert into 7.3M shares at $48.24 and the 4.75% notes into 7.0M shares at $99.66, about 36% more than the 39.4M now outstanding if both convert.³ A higher share price brings that dilution, and a lower one brings a share sale at a price that hurts existing holders. In August holders agreed to lift the limits on secured debt ahead of the 6.00% notes.²⁴
• Currency. The company reports in US dollars, but its operating subsidiaries and many of its costs are in Australian dollars. It booked $11.1M of unrealized foreign exchange losses in the half.³
• Export controls and regulation. The NVIDIA 8-K lists export controls and permitting among the risks.⁸ I take this up below, with numbers.

Geopolitical risk, and what it means for the value

SharonAI is a New York company with its equipment in Australia and New Zealand, chips designed by NVIDIA in the US and made by TSMC in Taiwan, a large customer in India, and lenders and investors mostly in the US.²³⁶ It sold its Texas data center stake in January, so no US site is left.³ I checked for an Israel link: the 10-K, 10-Q and S-1 never mention Israel.²³⁵

Export controls and tariffs. The 10-K names US export controls and Taiwan chip-making concentration as risks, and the NVIDIA 8-K names export controls and permitting.²⁸ Australia is a US ally and I found no rule in force that bars these sales. In May 2025 the Commerce Department said it would rescind the Biden-era “AI diffusion” rule and issue a replacement later.³⁵ Commerce decides, not the company. I put a rule that stops or delays the NVIDIA fleet before mid-2027 at about 1 in 15E (estimate), judging from Australia’s ally status and the open replacement. Tariffs are likelier. A 25% US duty on certain advanced chips began on 15 January 2026, aimed at imports into the US, and the S-1 warns tariffs could raise supplier prices.³⁶⁵ I put a 10% rise in unpaid hardware costs at about 1 in 10E (estimate); prepayments may fix part of the price.

Money and counterparties. The $3.6B gap depends on US lenders and investors, and a war or credit scare raises their price. The first loan costs 9.95%, and my base uses 11%.⁶ ESDS pays in US dollars, but its ability to pay is the risk, and I put a failure for most of its term at about 1 in 5E (estimate).⁵ NVIDIA is supplier, customer and revenue partner, and a report that it paused some financing is second-hand.²⁹ A change in its policy would reach SharonAI three ways at once.

Each row reruns my base with one change, over two years; odds are my guesses. The price already needs the NVIDIA fleet to fill and carries no cushion for these events: it sits 373% above the base before any of them. The first row rises because the NVIDIA contract earns less than its hardware costs in my model, so losing it with no penalty removes a loss. The filings do not say what ending it would cost, and without it the fleet cannot fill, which the price needs.