NVIDIA's $4.9B Australian Bet

🌾 Welcome to StableBread's Newsletter!

On March 31, 2026, a company that booked $294K of quarterly revenue signed a single customer contract worth $1.25B. Six weeks later, it signed another worth $950M. Then on June 12, 2026, NVIDIA agreed to a six-year collaboration, supplying up to 40,000 of its newest GPUs.

That company is Sharon AI (Nasdaq: SHAZ), an Australian neocloud that rents NVIDIA GPU compute on five-year, take-or-pay contracts. Management is guiding to a $470M revenue run rate by the end of 2026.

SHAZ listed on Nasdaq in February 2026 at $30.00, closed March at $22.73, and sits near $62/share today, with a market cap of ~$1.04B.

Sharon AI is unusual because it got called a fraud and funded by serious institutional money within weeks of each other.

On April 30, 2026, short seller Bleecker Street Research published a report calling the contracts "phantom" and dredging up self-dealing allegations against CEO James Manning from his last public company.

Three weeks later, Oaktree Capital Management closed a $350M convertible note it had agreed to days before the report and funded anyway. Notably, Leopold Aschenbrenner's Situational Awareness fund disclosed a stake the same week.

Below, I discuss which case the evidence supports and whether Sharon AI is an asymmetric play, given today's supply-constrained compute market and a business meaningfully more de-risked since its February 18, 2026 IPO.

Sharon AI buys NVIDIA GPUs, installs them in partner-owned data centers (listed below), wraps them in its own orchestration software, and rents the compute to enterprises, hyperscalers, and AI companies on multi-year contracts.

Manning’s framing on the Q1 call is that Sharon AI solves four scarcities at once: (1) compute, (2) power, (3) regulatory access, and (4) the HPC engineering talent.

On the compute side, the company is one of three NVIDIA Cloud Partners (NCPs) operating in Australia, alongside Firmus Technologies and ResetData. The certification requires building to NVIDIA’s reference architecture and comes with preferential access to new chips like the Blackwell-generation B200 and B300.

NVIDIA owns no part of the company. The 10-K originally described NVIDIA as "a strategic shareholder," and Sharon AI filed an 8-K on April 13, 2026 correcting the claim, stating NVIDIA holds no equity securities in the company.

The GPUs come from NVIDIA, the orchestration software is Sharon's own, and almost everything else runs through partners:

The customer pitch is sovereign AI infrastructure in a region short on compute.

Australia generates more energy than it uses, has close-ally access to the latest US chips, and is tightening data residency rules that keep data onshore, which feeds demand for exactly the local infrastructure Sharon sells.

With only 30M people, Australia can't absorb all the compute that surplus power could support. So the model is to build where the power and chips are, then serve the broader Asia-Pacific market, which sits within low-latency reach (81ms to Singapore, 130ms to Hong Kong).

The whole company runs on 25 employees, board members, advisors, and contractors as of March 18, 2026. That’s the partner model working as designed. It also means $2.2B of contracts rest on a 25-person team.

On June 12, Sharon AI signed a six-year strategic compute collaboration with NVIDIA itself. The stock jumped 25% on the news.

The deal enables 72 MW of new data center capacity in Australia, built to NVIDIA’s DSX AI factory design and scaling up to 40,000 Grace Blackwell GB300 GPUs. Underneath it sits a Master Cloud Services Agreement worth up to $4.88B over the term (per the 8-K).

The structure is what matters here. Sharon sells the NVIDIA-powered cloud services and owns the customer, while NVIDIA earns its usual hardware revenue plus a share of the cloud revenue on the supported capacity.

NVIDIA is also providing what the release calls "credit support." In plain terms, it lends its own creditworthiness to the financing, so a 25-person company can borrow against the GPUs and commit to capacity it couldn't fund on its own.

That goes straight at Sharon's funding gap, where ~$514M of secured capital sits against a $720M+ capex bill.

The deal also resets the capacity picture. Total AI factory capacity now reaches 132 MW, of which 102 MW is contracted to end customers, leaving 30 MW uncontracted. Management expects more than 55,000 NVIDIA GPUs deployed by mid-2027.

Two caveats before extrapolating.

Whatever NVIDIA takes comes out of Sharon's margin on that 72 MW, so it earns less there than the standalone unit economics show.

Second, that capacity only becomes take-or-pay revenue once end customers sign for it, which is exactly what the 30 MW of uncontracted capacity is waiting on.

Sharon AI didn't start as an AI company.

Over 2024 it acquired a set of older Australian storage operators, taking a controlling stake in Distributed Storage Solutions, a Filecoin business founded in 2021, then began repurposing the platform for GPU compute, buying 192 NVIDIA L40S GPUs and then 160 H100s.

In December 2024, NVIDIA certified Sharon AI as a Cloud Partner, which Manning said (at the March 2026 RedChip investor conference) reflected the largest installed base of GPUs for public cloud consumption in Australia at the time.

In March 2025 the company began building its Supercluster inside NEXTDC's M3 facility, a 1,016-GPU deployment on NVIDIA reference architecture that started on 16 H200s for testing.

By June 2025 it had fully wound down the legacy Filecoin operations. In January 2026 it announced the 1,000 B200 GPUs that fill out that Supercluster, built on Lenovo infrastructure and VAST storage and due online in the first half of 2026.

Then the company rebuilt its capital structure:

Within roughly 90 days, an Australian private storage operator became a Nasdaq-listed neocloud with $228M of new capital raised across the converts and the IPO ($103M + $125M, before costs).

One transaction funded much of this pivot.

In January 2025, Sharon AI formed a 50:50 joint venture with New Era Energy & Digital (NUAI) called Texas Critical Data Centers, to develop a 250 MW net-zero, gas-powered data center site in the Permian Basin.

Over 2025 the JV bought 438 contiguous acres, signed a non-binding LOI with a hyperscaler, and entered exclusivity in November.

Rather than fund the buildout, Sharon AI sold its 50% stake to New Era in January 2026 for ~$70M after adjustments, against ~$3.5M invested (per Manning at the March 2026 RedChip conference). Consideration was $10M cash, $10M of NUAI stock, and a $50M senior secured note at 10.0% interest.

New Era then repaid the note early. Sharon AI announced in an April 14, 2026 press release that it would receive the remaining consideration ahead of the note's June 30, 2026 maturity, and the Q1 2026 10-Q confirms full repayment in April.

The sale produced a $65.9M gain in Q1 2026, returning ~20x the invested capital in 12 months ($70M / $3.5M). Management has stepped back from powered-land development to focus on the cloud business, but it would do another deal like the Texas one if the right opportunity came along.

The gain is also a one-off, and it distorts the reported Q1 numbers, which I break down below.

Sharon AI signed four contracts in about four months:

That's more than $2.2B of total contract value (TCV), and on the Q1 call management guided to an exit-2026 revenue run rate of at least $470M.

These contracts are take-or-pay, meaning customers pay monthly for reserved capacity whether they use it or not. Management states this ensures 100% utilization recognition once deployed.

ESDS pays monthly in advance and must post $140M of letters of credit or bank guarantees as security. Renewal options are structured to be decided in year four.

Capacity has been upgraded twice to match. Sharon AI came to market in February guiding to 55 MW, moved to 70 MW with the ESDS signing, and on May 15 upgraded to 100 MW by early 2027, adding ~29.6 MW of new data center agreements (some in Australia, some in New Zealand).

The physical footprint behind that number spans NEXTDC's facilities (up to 54 MW, including the M3 Supercluster site and the 40 MW M2 block) plus a 15 MW, 120-month agreement signed in March with GreenSquareDC, which carries a ready-for-service target of September 26, 2026.

None of the take-or-pay book produces revenue today. ESDS revenue is expected to begin in September 2026, and the $950M contract by the end of Q3 and Q4 2026.

Between signing and billing sits 3-7 months of GPU delivery, installation, and commissioning, on hardware with 12-14 week lead times in a supply-constrained market.

The book is also concentrated. ESDS ($1.25B) and the unnamed customer ($950M) make up most of the $2.2B+ TCV, with only the undisclosed Canva and GMI deals on top.

Management provides the same unit economics across the deck and the Q1 call.

All-in capex runs ~$39M per MW (including non-recurring data center costs), and revenue runs ~$15M per MW at current pricing with an allowance for storage.

The March deck makes the illustrative case with a 4,608-GPU GB300 cluster (10 MW):

Storage is the upsell that makes the model stickier. NVIDIA's reference architecture suggests ~3 PiB of storage per 1,000-GPU cluster, and Manning said on the Q1 call that customers are requesting 6, 9, and 12 PiB, or 2-4x the reference amount.

Storage carries high margins, and management's argument is that accumulated customer data makes year-four renewal decisions much easier to win. Sharon AI has pre-purchased storage for ~70,000 GPUs' worth of deployments to get ahead of price rises.

These are management's assumptions, not results. The $4.00/hr rate is a base case in a market management itself describes as priced for scarcity, the payback assumes contracts stay whole for five years, and the six-year useful life will eventually meet NVIDIA's next architecture.

Sharon AI reported Q1 2026 on May 15. The headline net loss was $20.0M on $294K of revenue, but almost none of that loss came from operations.

The quarter contained two large offsetting non-operating items: (1) the $65.9M gain on the TCDC sale, and (2) a $70.2M non-cash loss from remeasuring the pre-IPO convertible notes, which are carried at fair value and rose from $129.0M to $199.4M as the company's equity value climbed.

Add a $13.5M tax charge on the TCDC gain, a $1.5M markdown on the NUAI shares, and smaller warrant and FX items, and you get to the $20.0M loss.

Strip all of it out and adjusted EBITDA was $(2.2)M, versus $(0.1)M a year ago.

Underneath the non-operating noise, the operating quarter was mixed:

Trade and other receivables jumped from $750K to $5.0M in the quarter, which looks alarming against $294K of revenue.

But $4.5M of that is GST (Australian sales tax) recoverable on equipment prepayments, not unpaid customer bills. Actual trade receivables are $571K.

The working capital story that matters is the $42.5M of equipment and lease prepayments that appeared on the balance sheet this quarter, deposits with WWT and Lenovo for hardware that hasn't been delivered yet.

Cash is going out the door months before revenue can come in. That's the model, but it makes the next two quarters' execution non-negotiable.

The 10-Q discloses that the ESDS contract alone requires ~$720M of capex for equipment and infrastructure. At management's $39M/MW, the full 100 MW buildout implies roughly $3.9B of capex over time.

Against that, the secured sources of capital:

And the announced-but-not-definitive sources:

On the Q1 call, Manning said the company is "funded fundamentally from all the equity components for all the currently announced contracts" between the Oaktree note, IPO proceeds, and customer deposits, and that it has hired investment bank Jarden to run a GPU debt financing process, with several term sheets in hand.

But the numbers don’t add up. Cash plus the new convert is ~$514M ($164M + $350M) against $720M for ESDS alone, before the $92M noncancelable WWT purchase order and the capacity behind the $950M contract.

Closing that gap depends on the Digital Alpha and USD.AI announcements becoming signed documents, the Jarden debt process landing, or more equity-linked paper.

The June 12 NVIDIA collaboration adds a fourth path, with NVIDIA providing credit support on the new 72 MW specifically so Sharon AI doesn't have to fund it the traditional way.

All this financing dilutes shareholders. The pre-IPO convertibles sit on the balance sheet at $199.4M fair value and convert into 8.25M shares. Their conversion trigger relates to events including a potential additional ASX listing, so they didn't convert in the Nasdaq IPO and keep generating non-cash losses as the stock rises.

The new $350M notes convert into another ~7.3M shares ($350M / $48.24), plus 1.25M options and RSUs and 0.44M warrants. On June 5, Sharon AI registered up to 11.3M shares from those notes for resale, clearing the way for the holders to sell into the market, the overhang that drove the June 13 pullback.

Stack it all up and the fully converted share count is ~34M against 16.7M outstanding today, roughly double. At $62/share that's a ~$2.1B fully diluted market cap, which is the number I'd use for valuation.

Three co-founders control the company. Sharon AI's dual-class structure gives each of the 136,341 Class B Super Common shares 160 votes, and all of the Class B sits with James Manning, Andrew Leece, and Nick Hughes-Jones (45,447 shares each).

Officers and directors as a group control 68.2% of the voting power while holding ~27.5% of the Class A shares.

Entities the founders control bought into the December 2025 convertible notes on the same terms as outside investors, and the founders' pay shifted from consulting contracts to standard employment agreements.

Both are disclosed and benign on their own, but they establish a related-party pattern that matters once the self-dealing allegations come up.

The board also separated the Chairman and CEO roles, appointing Andrew Penn as Chairman (May 21, 2026). Those are steps toward normal public-company governance, not away from it.

The disclosure record is less reassuring. Sharon AI reported a material weakness in internal controls over financial reporting related to "the accounting for complex financial instruments and transactions," identified at December 31, 2025 and not yet remediated as of March 31, 2026.

The Q1 2026 10-Q states plainly that disclosure controls and procedures were not effective. That weakness sits on the most consequential lines in the P&L.

The $199.4M convertible notes are valued with unobservable "Level 3" inputs and swung results by $70.2M this quarter, and the TCDC sale, warrants, and NUAI shares all run through fair-value accounting as well.

In other words, the company is telling you its controls are weakest exactly where its reported numbers come from.

The NVIDIA misstatement fits the pattern. The 10-K called NVIDIA a strategic shareholder, and the company retracted it by 8-K two weeks later.

A company that misstates its most important relationship and still hasn't closed its controls gap hasn't earned the benefit of the doubt, so I'd treat the fair-value lines as estimates until the material weakness clears.

On April 30, 2026, Bleecker Street Research published a short report titled "The CEO Mawson Accused of Self-Dealing Now Runs a Neocloud Built on Phantom Contracts and Questionable Financing."

It makes two charges: (1) the contracts are phantom, and (2) the CEO has a self-dealing history.

The self-dealing charge is the one that touches Sharon AI's own filings. Manning founded and ran Mawson Infrastructure Group (MIGI), a Nasdaq-listed bitcoin miner, until May 2023, and is still Managing Director of Vertua Limited.

Mawson is now suing him, alleging self-dealing that included more than $11M of payments to Flynt ICS, a shipping company he allegedly controlled through Vertua without disclosing his interest to Mawson's board.

That same Flynt shows up in Sharon AI's related-party note: the company paid Flynt, a Vertua subsidiary commonly owned by Manning, $167,638 in 2024 and $92,722 in 2025 for storage services. And several Mawson alumni now hold senior roles at Sharon AI, including CFO Tim Broadfoot, who was Mawson's CCO from 2020 to 2024.

However, the lawsuit targets Manning personally and is untested in court. Sharon AI's own 10-Q states the company is not party to any material legal proceedings. The Mawson team also built real infrastructure, over 300 MW across North America, with former sites now owned by Bitfarms and CleanSpark.

The market's answer is what makes the setup unusual. Oaktree had backed the February IPO, then struck its $350M convertible note days before the report and closed it three weeks after, with the short thesis fully public and the option to walk. That's doubling down, not a first bet.

Situational Awareness LP, Leopold Aschenbrenner's $13.68B fund, disclosed a ~796,000-share position (~4.8%) in its Q1 13F, built near the post-IPO lows.

Here's where I come out. The two charges don't carry equal weight…