IREN 2027 회계연도 1분기 실적 발표: AI 클라우드 매출 3배 급증, ARR 목표 1월로 변경

핵심 요약

• IREN의 2027 회계연도 1분기 매출은 전 분기 대비 66% 증가한 2억 2,800만 달러를 기록했다. AI 클라우드 매출은 183% 급증한 2억 달러에 달한 반면, GPU 설치를 위해 채굴 하드웨어 가동을 중단함에 따라 비트코인 채굴 매출은 2,800만 달러로 감소했다.
• 운영 연간 반복 매출(ARR)은 10억 달러에 달했다. 경영진은 현재 2027년 1월까지 운영 ARR 40억 달러 이상 달성을 목표로 하고 있으며, 공급업체 일정에 따라 2월 말까지 연장될 가능성도 열어두고 있다.
• 조정 EBITDA는 6,600만 달러로 증가해 전 분기의 14% 대비 29%의 마진율을 기록했다. 감가상각비 차감 전 AI 클라우드 매출총이익률은 86%를 유지했다.
• 매출 잔고(백로그)는 200억 달러를 넘어섰으며, 이 중 3분의 2는 투자등급 고객과 연계되어 있다. 최근 체결된 계약에는 메가와트당 2,200만 달러 규모의 5년 계약이 포함되었으며, 해당 계약의 고객 선급금은 관련 GPU 설비투자의 약 90%를 충당하는 수준이다.
• IREN은 9월 말 기준 사용제한현금 23억 달러를 포함해 58억 달러의 현금을 보유하고 있으며, 성장에 투입 가능한 현금 및 미인출 자금 조달 약정액은 약 90억 달러에 달한다.
• 경영진은 금융시장의 선별적 기조에 대응해 계획된 설비투자 속도를 늦췄으나, 2027년 말까지 0.8기가와트로 확장한다는 목표는 그대로 유지했다.

핵심 재무 데이터

사업 및 운영 실적

AI 클라우드가 주요 성장 동력으로 작용해 해당 분기 총매출 2억 2,800만 달러 중 2억 달러를 차지했다. AI 클라우드 비중 확대로 감가상각비 차감 전 연결 매출총이익률은 80%로 상승했다.

IREN은 비트코인 채굴에서 AI 인프라로 용량을 지속해서 전환했다. 경영진은 12월 분기 말까지 채굴 하드웨어의 가동 중단을 실질적으로 마무리하는 것을 목표로 하고 있다.

회사는 엔비디아 모범 클라우드(NVIDIA Exemplar Cloud) 지위를 획득한 '호라이즌 1(Horizon 1)'을 마이크로소프트에 인도했다. 경영진에 따르면 해당 부지는 토목 공사 단계에서 검증된 GB300 클러스터 구축까지 약 14개월이 소요됐다. 호라이즌 2~4는 완공 단계에 근접한 것으로 설명됐다.

운영 ARR 40억 달러 이상 달성 시점은 마이크로소프트가 요청한 네트워크 설계 변경과 공급업체 일정으로 인해 연말에서 2027년 1월로 미뤄졌다. 경영진은 네트워크 업그레이드로 인한 추가 비용이 전체 클러스터 투자 규모에 비해 미미한 수준이라고 밝혔다.

IREN의 계약 잔고는 200억 달러를 넘어섰으며, 이 중 3분의 2는 투자등급 거래 상대방과 체결된 것이다. 최근 영업 활동에는 5년 만기 기준 메가와트당 2,200만 달러 규모의 계약과 단기 계약 기준 메가와트당 4,000만 달러를 웃도는 계약이 포함됐다. 이전 세대인 H100 용량은 기존 가격의 두 배 수준에서 다년 계약으로 재계약됐다.

회사의 인프라 포트폴리오는 북미, 유럽, 아시아·태평양 전역에 걸쳐 3만 에이커 이상의 부지를 포함하며, 5기가와트 이상의 확보된 전력과 35기가와트 이상의 개발 중인 전력을 지원한다. 다만 경영진은 다수의 개발 프로젝트가 여전히 초기 단계에 머물러 있어 전력망 연계 계약으로 이어지지 않을 수도 있다고 경고했다.

경영진 가이던스

IREN은 운영 ARR 10억 달러를 확보한 상태로 12월 분기에 진입했으며, 이 중 4분의 1을 2분기 매출로 인식할 것으로 예상하고 있다. 경영진은 호라이즌 및 매켄지(Mackenzie)의 추가 증설 용량 대부분이 12월 말이나 2027년 초에 도입될 예정이어서 실질적인 매출 기여는 크지 않을 것으로 보고 있다.

판관비와 연구개발(R&D) 비용을 포함한 현금성 영업비용은 12월 분기에 전 분기 대비 약 5,000만~6,000만 달러 증가할 것으로 예상된다.

경영진은 2027년 1월까지 운영 ARR 40억 달러 이상을 달성한다는 목표다. 공급업체 납품이 지연될 경우 이 목표 시점은 2월 말로 미뤄질 수 있다. 해당 ARR 목표를 뒷받침하는 설비 용량에 대한 자금 조달은 완료된 상태다.

이번 분기 집행된 약 50억 달러에 이어 2027 회계연도 잔여 기간 동안 계획된 설비투자는 약 150억 달러다. 약 90억 달러의 현금 및 약정된 자금 조달원을 보유한 IREN은 추가로 60억 달러 규모의 계획된 설비투자에 필요한 자금을 마련해야 한다. 경영진은 전환사채나 유상증자를 제외하고도 잠재적인 자금 조달원으로 120억~150억 달러 수준의 가이드라인을 제시했다.

회사는 2027년 말까지 0.8기가와트 용량 확보라는 목표를 지속해서 유지하고 있다. 경영진은 신규 계약, 설비 시운전, 고객 승인에 따라 규모와 시기가 달라질 수 있다는 점을 들어 40억 달러를 초과하는 구체적인 ARR 목표는 제시하지 않았다.

위험 요인 및 주요 관전 포인트

• 공급업체의 납품 지연으로 인해 운영 ARR 40억 달러 달성 목표 시점이 2027년 1월에서 2월 말로 연기될 수 있다.
• 경영진은 자본시장이 더욱 선별적으로 변했다고 평가하며, 자금 조달 가능성과 비용이 자본 투입 속도에 영향을 미치고 있다고 밝혔다.
• 데이터센터 및 GPU 구축에는 네트워킹, 광케이블 포설, 클러스터 테스트 및 고객 승인 등 복잡한 시운전 과정이 수반된다.
• 업계 공급 제약 요인으로는 SSD, RAM 및 기타 부품 등이 꼽힌다. 경영진은 하드웨어 가격 상승 압박이 지속될 것으로 예상하면서도, 지금까지는 고객 가격 정책 덕분에 투자 회수 기간을 유지할 수 있었다고 설명했다.
• 35기가와트 규모의 개발 파이프라인은 여전히 잠재적 단계에 있으며, 경영진은 모든 초기 단계 프로젝트가 전력망 연계 계약을 확보할 수 있는 것은 아니라고 밝혔다.
• 회사는 2027년에 걸쳐 칠드레스(Childress) 공랭식 인프라 구축 속도를 조절하고 있으며, 내년 설비 가동에 필요한 모든 GPU 구매를 확정하지 않음으로써 시장이나 하드웨어 환경 변화에 유연하게 대처할 수 있는 여지를 남겨두었다.

애널리스트 Q&A 주요 내용

마이크로소프트 인도 일정: 경영진은 고객의 요청에 따른 네트워크 재구성 및 공급업체 일정을 지연 사유로 들었다. 호라이즌 2는 건설을 완료하고 시운전 중이며, 호라이즌 3은 현장에 모든 GPU를 배치하고 홀 전력 공급을 마쳤으며, 호라이즌 4의 GPU 인도는 수주일 내에 이뤄질 것으로 예상된다고 설명했다.

자금 조달 및 계약 구조: 경영진은 투자등급 고객과 일반적으로 더 많은 선급금을 제공하는 낮은 신용도의 거래 상대방을 결합한 포트폴리오를 강조했다. 최근 일부 고객 선급금은 관련 GPU 설비투자의 최대 90%를 충당했다.

설비투자 유연성: IREN은 2027 회계연도 지출 계획 축소가 고객 수요 둔화보다는 금융시장 여건을 반영한 것이라고 밝혔다. 토지와 전력망 연계, 데이터센터를 자체 보유하고 있기 때문에 장기 코로케이션 임차 의무 부담 없이 건설 일정을 조정할 수 있다는 것이 경영진의 설명이다.

GPU 수익성: 회사 측이 제시한 사례에 따르면 메가와트당 2,000만 달러의 수익을 창출하고 GPU 비용의 50%를 충당하는 선급금을 수취하는 5년 계약의 경우, GPU 투자 회수 기간은 18개월, 데이터센터를 포함한 회수 기간은 약 3년이 될 수 있다. 다만 경영진은 실제 수익률은 고객, 하드웨어 및 계약 구조에 따라 달라진다고 강조했다.

인프라 효율성: IREN은 정격 전력 대비 GPU를 초과 할당(오버서브스크립션)하고 최대 사용 시 자동 스로틀링을 적용하기 위해 엔비디아의 MaxLPS 소프트웨어를 평가 중이다. 아울러 냉각 대신 연산에 더 많은 전력망 용량을 배정하기 위해 더 높은 운영 온도를 검토하고 있다.

모듈러 공법: 경영진은 공사 기간을 단축하고 현장 인력 수요를 줄이기 위해 전담팀이 표준화된 모듈형 설계 및 외부 공장 제작(오프사이트 제조)을 추진하고 있다고 밝혔다.

실적 발표 컨퍼런스콜 전문

전체 실적 발표 컨퍼런스 콜 녹취록

경영진 발표

Operator

Hello, and welcome to IREN's First Quarter Fiscal Year 2027 Results Conference Call. [Operator Instructions]

I would now like to turn the call over to Mike Power, Vice President of Investor Relations. You may begin.

Mike Power

Thank you, operator. Good afternoon, and welcome to Ion's Q1 Fiscal 2027 Results Presentation. I'm Mike Power, Vice President of Investor Relations. And with me on the call today are Daniel Roberts, Co-Founder and Co-CEO; Anthony Lewis, CFO; and Kent Draper, Chief Commercial Officer. Before we begin, please note that this call is being webcast with an accompanying presentation. For those dialed in by phone, you can elect to ask a question through the moderator after our prepared remarks.

I would like to remind everyone that certain statements made during this call may constitute forward-looking statements. These statements are based on current expectations and assumptions and are subject to risks and uncertainties that could cause actual results to differ materially. Please refer to Slide 2 of the accompanying presentation on the Investor Relations section of our website and our SEC filings for more information.

During this call, we will present both GAAP and certain non-GAAP financial measures. Reconciliations of these measures to the most directly comparable GAAP measures are included in the appendix to today's earnings press release and the appendix to the accompanying presentation, both posted on the Investor Relations section of our website.

With that, I will turn the call over to Dan Roberts.

Daniel Roberts

Thanks, Mike. Good afternoon, everyone, and thank you for joining us. During the quarter, we signed new multiyear contracts, including with a leading frontier AI lab, taking revenue backlog to over $20 billion, 2/3 of it with investment-grade customers. We closed GPU financing, that with customer prepayments, funds more than 100% of the GPU CapEx for the Mackenzie deployments. We handed over Horizon 1 to Microsoft with NVIDIA Exemplar Cloud status. And finally, we advanced construction of new data center campuses across the globe, including our flagship 2-gigawatt Sweetwater hub. Compared to the last quarter, AI cloud revenue nearly tripled to $200 million. Operating ARR doubled to $1 billion, and adjusted EBITDA margin went from 14% to 29%.

One date has moved, so let me deal with it first. In July, we said more than $4 billion of operating ARR was targeted to be online by year-end. It's now January. Two reasons: one, a customer asked for a design change, which we agreed to, and it sits outside the original scope; and two, supplier timing. The data centers are practically done, and we're waiting for remaining equipment to arrive to finish commissioning. If suppliers deliver, we're targeting January. If they're late, as suppliers have been across the industry this year, it may run to the end of February.

Building at this scale is complex and even more complex when you own every layer: land and power, data centers, compute and the people to bring it together. We've spent years building all of it. Today, we run 3 businesses, summarized on Slide 4 of our presentation.

First, AI cloud, $1 billion of operating ARR today, more than $4 billion contracted and fully funded, ramping as the remaining capacity is commissioned. Second, the data centers, $6 billion of data center assets today, $10 billion targeted by June next year, 20-year assets that can carry several generations of compute, cash funded to date and unencumbered. Third and final, land and power, more than 30,000 acres across global markets, supporting both over 5 gigawatts of secured power and more than 35 gigawatts in development.

Many of those are early stage, and not all of them will convert to grid connection agreements. Securing land and power ahead of demand is what we've done since day 1, and it's why the platform today spans 3 continents. Most companies do 1 of those 3. We do all 3, which keeps more of the economics with us, gives us more ways to fund growth and more control over delivery.

Demand for compute hasn't slowed. Since quarter end, we've signed contracts at $22 million of annualized revenue per megawatt of IT load on a 5-year term and in excess of $40 million per megawatt on a shorter duration contract. Importantly, the 5-year agreement also includes a customer prepayment covering approximately 90% of the associated GPU CapEx. We've also expanded our AI cloud contract further with the same leading AI frontier lab and contracted H100s at double the original pricing. That's the pricing backdrop for contracting our Vera Rubin capacity.

Capital markets have become more selective in the last few weeks. And the questions have undoubtedly moved from how big the contract is to whether it can be delivered, financed and run at a profit. Those are the right questions, and this quarter for us answers them. Horizon 1 went from civils to an NVIDIA Exemplar validated GB300 cluster in about 14 months. JLL puts the global average at 18 months to build a 50-megawatt data center, and that's the building alone. We delivered the building and the cluster inside of 4 months faster than that benchmark. I'll come back to what that last stage involves.

Operating capacity rose 127% in the quarter to 0.1 gigawatts. We raised $19 billion in the 12 months to September, including $6.5 billion of GPU financing since May at 6% to 9.5%. Our AI cloud ran at an 86% gross margin, in line with last quarter.

Now this is where IREN differs from a typical neocloud. We don't lease someone else's data center and carry the liability on our balance sheet. We secure the power, we develop the land, we build the infrastructure and we deploy the compute. Each deployment leaves us with long-term assets we own.

Owning the power, the land and the data centers also means we set the pace. We want to build as fast as demand allows, but we're not obligated to develop every site at once. The power is secured, and the data centers are ours, so we don't sign contracts or raise capital to cover lease payments. Investment and contracting decisions are made on risk-adjusted returns, financing conditions included.

And the economics justify building. During the quarter, average revenue per megawatt on our 3- and 5-year contracts continued to increase, now supporting GPU paybacks of around 18 months or even lower depending on the level of customer prepayments. At Mackenzie, customer prepayments and GPU financing together covered more than 100% of the associated GPU CapEx. The contracts signed since quarter end reinforce that picture, attractive pricing alongside customers funding a substantial portion of the GPU CapEx.

And as we add capacity, we're strengthening how we deploy and how we operate it. Our work with NVIDIA and the Mirantis and [ Cordant ] AI teams is improving how we provision, manage and support these environments. Existing customers are renewing and expanding with us. New logos are proving demand, and the renewals are proving ongoing delivery.

Now to delivery on the ground. Slides 13 and 14 of the presentation show progress site by site on our targeted expansion to 0.8 gigawatts by the end of '27 and data center expansion beyond that. The initial 0.3 gigawatts is in the final stages, and construction for further capacity is underway across 6 sites.

That last mile is the most complex. It's the part you don't see from the drone photos. GPUs wrapped, miles of fiber pulled, whole clusters connected and tested and the data center business handing over to the AI cloud business. The expertise behind these deployments is also shaping the next generation of AI factories. Led by our Chief Innovation Officer, John Gross, we're working closely with NVIDIA to advance data center design and develop even faster ways to build.

That work also means getting more out of the sites we already own, using power headroom within existing connections to support more compute. Mackenzie is the example, 100 megawatts of power secured, mass grading now started for new liquid cooled data centers, built to the same design we've already delivered.

Slide 17 shows the pipeline behind the data center and AI cloud businesses I just described. Our development team keeps originating sites ahead of the next phase of growth. Those more than 30,000 acres span North America, Europe and Asia Pacific. That gives us choices about where we serve customers and where we progress sites as local conditions allow.

We don't need every market to move at the same pace. Where we build is just as important as how we build. We focus on rural locations with appropriate setbacks from neighbors and engage with communities and regulators before progressing grid applications to understand local priorities from the outset. Our recently announced 900-megawatt Redbank site is a good example of this in Australia.

We hire locally, we work with local suppliers, and we prioritize bringing water to site for our closed-loop cooling systems to reduce reliance on local suppliers. We want our projects to bring lasting jobs and investment to these communities while respecting the needs of the people who live there. We're being deliberate about how we grow, contracts with customers who renew, returns that justify the capital and the financing attached to each deployment.

With that, I'll hand over to Anthony to walk through the numbers and how we're funding the next phase.

Anthony Lewis

Thanks, Dan, and good evening, everyone, and thank you for your attendance. Starting with our results, we delivered another quarter of significant growth in AI cloud revenue, supported by continued execution. AI cloud revenue was $200 million for the September quarter, up 183% quarter-over-quarter, reflecting the continued deployment and monetization of our data center and GPU capacity. Total revenue was $228 million, up 66% quarter-over-quarter, with Bitcoin mining revenue falling to $28 million as we continue to actively decommission Bitcoin mining hardware ahead of GPU installations. We aim to substantially complete this transition by the end of this quarter.

Gross margin before depreciation and amortization was 80%, up from 76% in the prior quarter as AI cloud revenue represented a larger share of total revenue. AI cloud gross margin before depreciation and amortization was 86%, consistent with the prior quarter. Adjusted EBITDA was $66 million, representing a margin of approximately 29% compared with 14% in the prior quarter. Operating expense, including SG&A and R&D, was $224 million, up $96 million quarter-over-quarter, reflecting an increase in cash operating expense of $33 million as we added the Mirantis operations from the middle of the quarter and continue to significantly expand headcount to support our growth and after stock-based compensation of $63 million.

After taking into account other items, including depreciation, amortization, financing costs and other expenses and income, that gave a net loss for the quarter of $263 million. The results demonstrate our ability to convert our infrastructure investments and customer commitments into revenue-generating operating capacity. As Dan mentioned, our 0.3 gigawatt expansion is nearing completion, with the associated ramp to more than $4 billion of operating ARR expected to extend into early 2027, primarily due to customer-requested design changes and supplier timing.

We expect ARR to grow meaningfully beyond that level as additional capacity comes online through calendar year 2027. However, we're not providing a specific target at this stage as both the level of ARR and its timing will depend on the contracts that we sign and when capacity comes online and deployments are accepted.

Now looking towards the December quarter specifically, we exited September with $1 billion of operating ARR, 1/4 of which we expect to recognize as revenue during the second quarter. We do not expect material revenue contribution from incremental Horizon and Mackenzie capacity, most of which will arrive late in December or early calendar year '27. For the December quarter, we expect cash operating expense, including SG&A and R&D, to increase approximately $50 million to $60 million sequentially.

Before turning to balance sheet and funding, I'd like to spend a few minutes walking through recent contracting activity and how prepayments support our funding plan and generate enhanced paybacks and returns on our investments. As Dan covered, demand for our AI cloud services remains strong, and we're seeing attractive opportunities across hyperscalers, frontier AI labs and other customers. During the quarter, 3-year pricing was at $19 million to $23 million per megawatt of IT load and 5-year pricing at approximately $17 million. And as Dan noted, after the quarter end, we have signed contracts at $22 million to $40 million per megawatt. This demand backdrop gives us the ability to be selective about the contracts we sign and how we structure them.

Customer prepayments are an important part of our model. As Dan mentioned, some recent contracts have included customer prepayments representing up to 90% of associated GPU CapEx. Prepayments reduce upfront funding requirements and help align capital deployment with contracted demand. Combined with equipment financing, they can substantially reduce the amount of corporate capital required for deployment and reduce the cash payback period on total investment.

The table on Slide 9 of the accompanying presentation illustrates this point for a hypothetical 1-megawatt liquid cooled deployment for around $60 million of CapEx, 1/3 for the data center and 2/3 for GPUs, and customer prepayments of 0% to 90% of the GPU costs. On a 5-year contract of $20 million per megawatt and with the prepayment covering 50% of GPU CapEx, we get a payback on the GPUs of 18 months. And if we include the cost of the data centers, the payback is around 3 years.

These are illustrative, and deal economics and payback periods will ultimately vary by customer, hardware and contract structure. But it shows that our focus is not simply on how much revenue we can track, it's on how efficiently we can fund that revenue and the returns it generates.

Now turning to the strength of the balance sheet. At September 30, we had $20.6 billion of total assets, including $12.3 billion of property, plant and equipment. That includes $6 billion of unencumbered data center assets. Cash totaled $5.8 billion, including $2.3 billion of restricted cash set aside primarily to finance GPU deployments. Total debt was $10.2 billion, largely in low coupon convertible notes and with maturities concentrated in 2032 and beyond.

Over the 12 months to September 30, we secured $19 billion of funding across a range of sources, including customer prepayments and secured GPU financing. Taking into account the balance of this financing, which is yet to be drawn down, we have a total of approximately $9 billion of cash and undrawn commitments to fund growth.

In terms of outlook for CapEx, we spent a total of approximately $5 billion in the quarter. Our planned CapEx for the remainder of the fiscal year is approximately $15 billion. This is down from the prior estimate, reflecting the flexibility to pace data center and GPU CapEx. So with cash and committed funding sources of $9 billion, that leaves $6 billion additional planned CapEx to be funded.

With this updated plan, we expect to see deployments of Childress's air cooled capacity spread across calendar year 2027, which were previously expected to be more front-loaded. We will target additional investments in support of longer-term capacity targets, and we'll continue to adjust our plan based on a range of factors, including the fundraising environment, contracting and customer requirements and other factors. As Dan noted, that's the strength of our vertically integrated model, which gives us maximum flexibility to adjust our planning to different conditions.

Now turning to the funding plan. As shown on Slide 19 of the presentation, with $10 billion of unencumbered data center assets projected at the end of FY '27 and $6 billion of unencumbered GPUs before further secured financing and cash flows supported by multiyear contracts with high-quality customers, 2/3 of our backlog being Microsoft and NVIDIA as investment-grade counterparties, that gives us substantial funding options across data center financing, GPU prepayments and financing and other corporate-level initiatives. Indicatively, $12 billion to $15 billion is against the $6 billion to be funded. And that's in addition to other capital market sources such as convertibles and equity, which leaves us very well placed to fund our fiscal year '27 growth and beyond.

Touching on a few aspects of our plans more specifically regarding our data center financing priorities. As you would expect, a near-term focus will be Horizons 1 to 4, which service the Microsoft contract. We continue to advance a number of parallel initiatives, taking into account its advanced stage of construction and delivery, the underlying AAA-rated cash flows for the Microsoft contract and the high-quality long-term physical infrastructure.

From a GPU perspective, as previously discussed, prepayments will continue to be a core pillar of our overall funding strategy. Across the portfolio, we are targeting total financing for GPU CapEx, including prepayments, of 90% of associated GPU CapEx. This is consistent with our previously announced investment-grade and noninvestment-grade GPU financings, where we secured funding of circa [ 95% ] and greater than 100%, respectively.

In terms of our near-term priorities for GPU financing, we have advanced work streams underway for the Childress air cooled deployments. These will be anchored by the NVIDIA and leading lab contracts, the latter of which was recently expanded. In addition to data center and GPU financing activities, you can expect other corporate level initiatives over time, and we will continue to maintain an appropriate balance of debt and equity funding as the platform scales.

In closing, this quarter showed our ability to convert infrastructure investment and customer commitments into operating capacity and high-margin revenue, and we feel very well placed to continue to scale the business with the ability to tailor our growth to external conditions. With that, I'll turn it back to Dan.

Daniel Roberts

Thanks, Anthony. So where I started on delivery. We're delivering across 3 layers of our business: securing land and power, building data centers and bringing AI cloud capacity into service. By June next year, we expect to have $25 billion of data center and GPU assets. We raised $19 billion in the 12 months to September. The capacity behind the more than $4 billion of ARR is fully funded. The additional $6 billion of planned expansion CapEx sits against $12 billion to $15 billion of sources before converts or equity.

We also have the flexibility to pace or accelerate that investment. We'll keep building the contracts, funding, and returns justify it. The next milestone for us is the more than $4 billion of operating ARR. It's funded, and it's targeted for January.

Thank you for joining us. We're happy to take your questions.

Operator

[Operator Instructions] Our first question comes from the line of Paul Golding with Macquarie.

질의응답

Paul Golding

Congrats on all the progress and the new leasing. I wanted to ask, there's been a lot of press around the delivery of the Microsoft capacity. And I was wondering, Dan, if you could give any additional color there on some of the dynamics you discussed in the prepared remarks, maybe parse out how much of the supplier delay related to the change orders or if they're unrelated? And just any other clarity around some of that commentary we've been hearing in the marketplace?

Daniel Roberts

Thanks, Paul. Absolutely. So as we announced back in August, Horizon 1 was handed over to Microsoft. And subsequent to that, they made a request to reconfigure it, upgrade some of the network infrastructure. And the purpose of that was really to enable the 200 megawatts of capacity to service these latest generation frontier models.

And not everything -- this market moves very quickly, right? So signing the contract back in 2025, not everything was contemplated. So look, we said yes. Of course, no problems. We're here to say yes to customers, not no. As a result, there will be a small delay. We're targeting January, as we noted in the release. It does depend on supplier timing, and it's been a tough year for suppliers and reliability of delivery. So there is every chance that moves to the end of February.

But in terms of the construction progress, we're in a good spot, and Horizon 2 is built. Power's on at the rack in all 3 halls, the GPUs are racked, the fiber's in, deal is actually commissioning now. So there's 1 integrated test there. Horizon 3 has got all the GPUs on site, all 3 halls energized. Horizon 4's GPUs are delivering in the coming weeks.

So yes, we have noticed the speculation, but it's a great relationship. I think in some ways, it's a testament to the relationship and the business model that we were able to pivot and be flexible around this network infrastructure. So I'll give you one example. If we were a different type of business where we're an asset-light neocloud with a colocation partner, I'm not sure Microsoft would have made the same request because as a cloud provider, you would then need to go back to your colocation provider and ask for those changes.

So look, at 31 December, would we love to hit it? Absolutely. Very close, but it's looking like it will be January, potentially February, depending on supply performance.

Paul Golding

I just wanted to follow up just on those points, if I may. What -- how should we think about the CapEx per megawatt then as it relates to maybe some of this upgraded networking gear? Is that equation changing at all in terms of that 14 to 16? Or is this something that we might expect to be more pervasive across the leasing that's already been done in terms of change orders? How should we think about that? And any -- the tenant responsibility to cover any change orders along those lines?

Daniel Roberts

So the CapEx -- the incremental CapEx associated with the network upgrade is entirely immaterial against the backdrop of the CapEx for the broader cluster. We're talking new cables, labor and just a different methodology in terms of implementation. So I wouldn't get caught up on that.

In terms of what it gives us is it's effectively upgrading the full 200 megawatts for those latest-generation cutting-edge use cases. And also importantly, I should have mentioned, we're then compounding that learning into all our other sites under construction. So I guess this is a benefit of kind of being at that leading edge delivering. We've always said you need to be doing things to learn how to do things in a way. So the more you do, the more you learn, the more those learnings compound and roll into the platform more broadly, which is what we're seeing at the moment.

Operator

Our next question comes from the line of Brett Knoblauch with Cantor Fitzgerald.

Brett Knoblauch

Maybe to start, there's been a lot of talk recently about kind of modular data centers. And I'm curious if you guys are looking into that at any one of your sites that span across the globe? I know NVIDIA is coming out with their own modular products. A peer, I think, [ is SunPower ] land lease for that product, if I had to guess. Is that something you're looking? Or do you guys want kind of your own design and build? Or are you flexible on that?

Daniel Roberts

This is something we've actually been looking into for quite some time. And we've got dedicated teams internally, specifically set up in the last 9 months who are dedicated to construction methodologies and making things quicker and easier. And central to that is a couple of key concepts. One is modularization and the ability to plug and play more components on site. So having that standardized methodology, standardized supply chain so that it does become more plug-and-play and moving a lot of the construction offsite into factories where it's more scalable.

So the modular design is something that we've been looking at for quite some time. We've got a very close relationship with NVIDIA, as you'll be well aware. We're talking to them about the different designs. Whether or not that specific modular design is implemented the way that has been presented is something we're working through, but certainly, the concepts and the overall North Star of trying to make construction simpler, quicker, easier, reduced labor requirements on site is absolutely where the industry and we are heading.

Brett Knoblauch

And then I guess if I could just follow up on how you guys are thinking about maybe contract duration? I know you guys are signing maybe a barbell approach, long-short pricing, and economics kind of vary between the both. But in this environment where funding is a concern, is having more IG-backed customers an emphasis for you? Or do you not view it that differently if you're signing deals with customers that are prepaying for the entirety of the CapEx in some cases.

Daniel Roberts

Good question. The answer is both. They're not mutually exclusive, and you alluded to the barbell approach. So we've got a $20 billion backlog of customer contracts at the moment. And it's been very deliberately constructed where 2/3 of that is investment-grade. So we're talking the likes of Microsoft, NVIDIA, $4 trillion credits standing behind that. And then, yes, there is a portion that is a lower credit, but they tend to go hand in hand with higher prepayments, as you mentioned, it is central to our broader strategic objective of servicing the broader ecosystem.

So we're very happy with the investment-grade counterparties. We've got a fantastic relationship with Microsoft, a really good relationship with NVIDIA, and we'd love to continue servicing their requirements. But equally in parallel, we've said for quite some time, it's important to be servicing the broader ecosystem.

So I feel like we've constructed that customer book deliberately and in a position that leaves us in a good spot today where we've got forecast $25 billion in data center and GPU assets at 30 June next year against that sort of credit profile, which dovetails into the financing and how we look to grow from here.

Operator

Our next question comes from the line of Joseph Vafi with Canaccord Genuity.

Joseph Vafi

Terrific results. Totally appreciate your comments, Dan, about the vertical integration here and your ability to dial up or dial down development and rollout. Just maybe against that backdrop, could you provide maybe some color on things that you look at in the market that may gate your development -- pace of development, maybe things that you look at that might result in you dialing it down or dialing up pace of development of supply chain financing, demand, ability to physically develop the infrastructure or other things?

Daniel Roberts

Thanks, Joe. So we've been chatting since our IPO 5 years ago, almost to the date, coming up. And I think what we've seen over that time in the 3 years of our history prior to that is we've continued to build up this asset base with a strong foundation for the long term. This is when there's never short-term vision. It was founded 8 years ago with a long-term vision around this digital exponential demand for compute.

So we thought it appropriate today to disclose a little bit more about that development portfolio. So we've got 30,000 acres of land now secured globally underpinning 5 gigawatts of secured grid connection and another 35 gigawatts of prospective grid connections. And every layer has been built on top of every preceding layer. So there's the land and power. We then have cash funded, every bit of our data centers to date, not an ounce of debt against it. So it's cash funded, $6 billion today, forecast to go to $10 billion by 30 June. And then we buy the GPUs on top of that.

And the benefit of this model is we aren't a WeWork business model in the sense we're not sitting in between a provider of infrastructure and a user of the compute that leverages that infrastructure. And because we own and control our own destiny. And I know these are words that you've heard many times over the last 5 years. We are in full control. We can flex up, we can flex down. And this goes to global diversification of projects, being able to throttle up and down depending on the local environment. It goes to availability of capital, it goes to strength of demand.

And then I think that's a really good segue just to comment on the market environment that we're seeing. It's actually incredible when we look at how the markets are behaving, but then look at what's happening at the customer level and the actual substance and fundamentals of our business model. But the demand -- I know we tend to say this each quarter, but I will reiterate it again. It has never been stronger.

And as we just announced, these are deals that were signed this week. So this isn't old information. We just had a customer agree to fund 90% of the GPU CapEx. We signed a short-term deal at $40 million per megawatt. We had one of the leading frontier AI labs, the same one we announced previously a couple of months ago, expand their capacity again. And also, I think it's worth noticing that we have some older generation H100s where we just recontracted those old units at double the original pricing.

So the demand for compute, the supply issues that we're all seeing is creating this exceptional opportunity as underlying business level. And the dislocation with financial markets is really stark at the moment. And I guess the benefit of our model is we can be patient, we can work through these market dynamics and be ready to flex up our growth once the financial markets recorrect.

Operator

[Operator Instructions] Our next question comes from the line of Greg Lewis with BTIG.

Gregory Lewis

I did have a question, and Daniel, I think you were talking about it, alluding to it with the flexibility around the spending. But kind of as we look at what -- where CapEx for fiscal year '27 stands today, it's come down from the previous quarter, which looks to be down from the quarter before that. So just kind of -- like as we think about that flexibility around the CapEx, what's kind of driving that? And if you could provide some more color around why that looks like it's ticking down for this year?

Daniel Roberts

Good question. And it absolutely is ticking down. We have deliberately flexed it down. And it's not because we want to. It's because -- and it's not because we need to because of a lack of growth opportunities. The demand is there. The customers are there. The sites, the 30,000 acres, the 5 gigawatts of secured power, all the ingredients are there. It's purely and simply financial markets in the current state. And capital needs to be respected in the current environment, and we're not going to grow at all costs. We're going to find that middle ground based on the availability of funding, the cost of funding to throttle that growth curve.

As we noted in the deck in our prepared remarks, we don't sign 20-year colocation agreements where we've got a 20-year liability. We don't -- we haven't signed $100 billion TCV contracts with AI labs. We haven't signed up to tens and tens of billions of GPU CapEx and created this business model where we need to go and raise money at all costs. In fact, that perception around a neocloud in inverted commerce, I feel like we're diametrically opposed in terms of our core underlying business model, which has been constructed literally from the ground up, the land, the power, the data centers, cash-funded.

And the strength and resilience of that platform allows us to ride the volatility in financial markets, but be poised to capitalize when it returns. And it's a pity because the demand signals are there. As I mentioned, all those data points are this week, like this is live. So what's happening at the business level is very disconnected from what we're seeing in the market.

Gregory Lewis

Yes. No, that's super helpful. And then, I mean, you alluded to the strength in GPU pricing, even for the H100, spot on. I'm curious, can you share any details around the contract durations and maybe some of these recontracting that you're seeing with some of your GPUs?

Kent Draper

Yes, I'm happy to jump in with some additional detail there. And obviously, we don't comment specifically on individual contracts, but...

Gregory Lewis

No, a broad range is fine.

Kent Draper

Yes. I mean, the reference that Dan gave on the H100s to doubling the price that they were previously being leased out, that is now for a multiyear term that, that additional pricing has been locked in for. So yes, it's not that we were able to double capacity by selling on an ultra short-term contract. This is genuine long-term contracting for those H100s that has still resulted in a doubling in price. And as Dan alluded to, I mean, that is just indicative of the strength of the market at the moment and the level of demand that there is.

Operator

[Operator Instructions] Our next question comes from the line of Michael Donovan with Compass Point.

Michael Donovan

Going back to your commentary on flexing. When looking at your sites, how are you thinking about prioritizing sites such as Oklahoma and Spain, if financing is the governor?

Daniel Roberts

So there's a portfolio of sites, obviously. Our philosophy and mentality has always been cheaper call options, without that sounding the wrong way. So it's about building optionality on top of a strong foundation. And what you find is through the value creation curve, all the way from development of land and power through to buying the computers, there is an inverse correlation between time and the amount of money you need to spend to preserve that time to compute that we keep talking about.

And to be a bit more specific around that, keeping projects ticking along is cheap. And then it starts to get more expensive. So signing a land option is relatively cheap, buying the land is relatively cheap. Building money on -- spending money on a grid connection then becomes a little bit more. Civils, land clearing, laying down concrete foundations. You can see how the CapEx starts to grow as your time to compute gets closer.

So what we've done throughout our history is make live time decisions around capital allocation where we preserve optionality around that. So construction activity is happening at 6 sites at the moment. That's likely to expand in the short term.

And then it comes to capital availability. And as I've mentioned a few times, we know what's happening in the financial markets, but there are a lot of options for us on the capital side even beyond what we've tried to outline in a lot of detail in our slide deck today. A lot of our background comes from private markets, real assets, infrastructure funds. The depth of capital generally in the world, whether it's listed or unlisted, I believe the fundamentals and what we're seeing in terms of customer demand will prevail, and we're having a lot of conversations around funding that.

My personal view is that this dislocation between the fundamentals in the market sentiment at the moment is unlikely to persist very long. We are prepared if it does. But there are a number of options that we can continue monetizing that 5 gigawatts, even the absence of equity market recovery.

Michael Donovan

Appreciate that. And you mentioned working with NVIDIA to increase the speed of site development. Are you also utilizing NVIDIA software to eke out more power? And if so, how -- what commentary can you give around efficiency gains and perhaps more GPUs being able to be deployed at a given site?

Kent Draper

Yes, I'm happy to jump in and give some additional color there. So one of the reasons why we are working so closely with NVIDIA is they are not just focused purely on the chip side of the business, but they are looking at the broader data center as the unit of compute. And you hear Jensen talking about this a lot. It's no longer just the chips or the servers. It's how that entire integrated data center unit functions to ultimately produce compute.

And there are a number of elements within that. Part of it is software, and we have been looking at some of the NVIDIA software, which is called MaxLPS, to enable more nameplate GPUs per megawatt than you would have if you just limit the GPUs to the exact power envelope. And the reason that you're able to do that is the vast majority of time, GPUs are actually operating under 100% of their nameplate power draw. And obviously, if you're just sizing to the nameplate power draw, that means you have stranded megawatts a lot of the time.

What their software allows you to do is actually oversubscribe the GPUs. And in the rare instances where they are all operating near nameplate capacity, the software will actually step in to make sure that you throttle the GPUs so you're not exceeding nameplate capacity. So that's one example where you're able to get additional GPUs per grid-connected megawatt.

Another example is NVIDIA is doing a lot of work around the temperatures at which you can operate. And to the extent that you're able to operate at a higher temperature, it means you can direct less of your grid-connected megawatts towards cooling infrastructure and more of your grid-connected megawatts to the actual underlying IT compute. So those are a couple of examples, but there are numerous other elements that we continue to refine and work on with NVIDIA.

Daniel Roberts

And I think just to add to those comments, there's also the physical infrastructure advantage as well, the fact that we own the grid connections, the data centers and the opportunity to optimize beyond the software envelope rather than being stuck in the middle between an infrastructure provider and the end customer. So a lot of this is why we have been asked to upgrade clusters because we can, and we have that execution capability. It's a lot of the reason why we are getting expansion contracts from leading AI frontier labs and getting a lot of this customer traction because they have confidence in our ability to execute because we own and control all the stack.

And I think that data point in the deck really sums it up. We delivered Horizon 1 data centers and GPUs. Our first-ever liquid called data center, we delivered it in 14 months versus an 18-month industry standard for just the data centers. So it's really hard to understate the value of that physical infrastructure. And I know we've been talking about it for a while, but I think this market, particularly on the customer side at least, is really starting to appreciate it.

Operator

Our next question comes from the line of Mike Colonnese with H.C. Wainwright.

Michael Colonnese

So just curious, what portion of the 500 megawatts of planned incremental data center capacity for '27 is now contracting for after the recent deals you signed here? And just curious how we should think about future contract signings relative to phased data center deployments next year as you look to fill out this remaining capacity? And how do you balance that with [indiscernible] back up? It would be great to get your thoughts there.

Daniel Roberts

So in terms of the 0.5 gigawatts, we haven't guided specifically on that, the exact composition. But within that, it's clear that NVIDIA is a proportion of that capacity. It's clear that another good credit with one of these leading AI frontier labs is another credit behind that. So we've got the air cooled at Childress that's rolling out in 2027. And then we've got [ VRs ] arriving later in the year, so the latest generation liquid for both Childress and Sweetwater 1.

And I think that's where we're particularly excited. We haven't mentioned on this call, but the supply of VRs next year is exceptionally scarce, and we're seeing a lot of interest, and that's something that's really exciting to us in terms of the prospective economics on it. But most of the near-term activity we're seeing is on the air cooled at Childress.

Operator

[Operator Instructions] Our next question comes from the line of Michael Ng with Goldman Sachs.

Michael Ng

I just have two, if I may. First, just on the change in the CapEx outlook. Could you just talk a little bit about that flexibility that you mentioned? And why does the kind of change in the CapEx outlook not change the expectation in terms of data center capacity kind of targeted for year-end 2027?

And then second, I just wanted to ask that in the event there's some sort of community opposition that results in a delay, could customers kind of exit any leases or get any of their kind of GPU deposits back if there were something that causes a substantial delay?

Daniel Roberts

Sorry, would you mind repeating the first half of that question? I've got the [ dirty ] community opposition.

Michael Ng

Yes. So the first part of the question was just on the reduction [ for this year ]. Yes. Why is -- how does the CapEx come down, but the kind of planned capacity kind of stay the same?

Daniel Roberts

I think the simplest answer to that is it's indicative of the investment in the 35 gigawatts of development that we're pursuing. We're not here for 200 megawatts of Horizon 1. We're not even here for the 5 gigawatts of secured power, where we've got a vision. We want to pursue it. We've got the ability to flex up and down the capital spend.

But it goes back to what I was saying earlier around long lead items and getting ahead of things. So starting fresh now and looking for large-scale power. if you can find it, you are into the 2030s very comfortably. So it's important we keep that 35 gigawatts ticking away. But equally in the presentation where the CapEx becomes more meaningful, the long-term development is not overly material.

You can see we've started building across many, many sites. And we haven't provided guidance for 2028, other than we've said that we're constructing capacity to be available then. So there is the opportunity to flex up and down the cadence of CapEx spend and the pace of growth across all of those sites.

And when you see the market turn the way it has in the last few months, the first thing we do is just pace that construction and pace that CapEx spend and just hold it off and wait for market conditions to turn. So this is the flexibility of the model when you own and control everything. And we're not building into 20-year colocation liabilities where we're building assets and where you get the choice when and how to build on top of those assets rather than be forced to double down when markets are conducive.

In terms of the community opposition, I mean, when we say opposition, like our view is what we're seeing is entirely sensible. AI isn't unknown. And I think data centers are the physical manifestation of AI and this big thing that's happening in the world. And I think whether it's Texas, whether it's Australia, whether it's Spain, everywhere, all we're seeing is people wanting to ensure that data centers pay their own way and do the right thing.

And it feels completely reasonable. We hire locally. We work with local suppliers. We prioritize bringing the limited water we need for our closed-loop systems from off-site. I think everyone wants these projects to bring lasting jobs and investment to these communities and respect the needs of the people that live there.

So I think we're just going through that period where the size and scale of the opportunity in the end market for AI, I think what we're seeing is a political and community level is indicative of the size of the market and the growth that we're growing into. And there's a sensible desire to just make sure that's done properly.

Michael Ng

Great. Thank you, Dan.

Kent Draper

The one other comment I'd add to that as well is we take into account these issues from day 1 of our development, including the siting of the projects themselves. So if you look across our fleet of projects, these are not data centers that are located in urban environments that have specific near-term impacts on residential areas with construction noise and operational noise. Our data centers are located in regional areas where the additional economic development that is brought in by these projects is enormously beneficial for those local communities. And we work very closely with those local communities all the way from, as Dan mentioned, local hiring, use of local contractors, to community grants programs and also in the case of new developments, community consultation and education. So yes, this is something that we literally build into our development process from day 1.

Operator

Our next question comes from the line of Jonathan Petersen with Jefferies.

Jonathan Petersen

Okay. Maybe first question, maybe I'll stick on the same topic. Curious if you could remind us at Sweetwater, I know you have conditional baseload approval. Is your expectation still that you get final interconnection approval in December? And then are there any other outstanding permits or just anything that we should think about in terms of bringing Sweetwater online?

Kent Draper

So just to be clear, at Sweetwater, so Sweetwater 1, which is our 1.4-gigawatt project, that is already connected to the grid. So I'm not quite sure what the December date was that you were referring to, but that project is already grid-connected. As we announced a couple of months ago, both Sweetwater 1 and Sweetwater 2 were included in the batch -- Batch 0 -- or sorry, the baseload of the batching process. So that means all of the capacity, the 2 gigawatts across those sites is secured.

In addition to that, we had a number of other projects that qualified for the Batch 0 process as well. As we mentioned at the time, we only announce projects once they actually sign binding grid interconnection agreements. And so we plan to announce those projects in due course, but we do have additional projects outside of Sweetwater that are included as well.

I'd also note for both Sweetwater 1 and Sweetwater 2, those projects are already well under construction, as Dan pointed out during the presentation. You can see very clearly the status of those, but there is a significant amount of construction work on site already underway, and we obtained the relevant permits throughout the construction process. So hopefully, that gives you a bit more color as to the status of each of those projects.

Jonathan Petersen

Yes, that's great. Sorry for mixing up the dates. Another question here. Can you talk to us about the cost of purchasing new GPUs and how that's trending? And do you see GPUs being available to you as long as you can bring data center infrastructure online? And I guess, how do you expect those costs of purchasing GPUs to trend in the coming quarters? I guess, in other words, what I'm trying to get at is, will the cost to purchase GPUs catch up to where you're leasing them at today? Or do you expect these very short payback periods to persist for a while?

Kent Draper

So I think it's well publicized that there continued to be supply challenges throughout the industry, in particular, SSDs and RAM. So these various forms of memory have been constrained for quite a while. There are other components that continue to be constrained. And in that environment, we would expect to continue to see GPU pricing increase over time. And obviously, that is driven by the end-user customer demand. So again, it's always interesting to see these dynamics where ultimately, the customer demand is just so strong that, that is what is causing price increases in terms of hardware.

What we see on the pricing side with customers is that any price increases flow through immediately to hourly pricing under contracts. And so the payback periods continue to at least remain where they are, and we've actually seen a shrinking of some of the payback periods over the past 6 months as that market demand continues to come through. So to answer your question in short, yes, the payback periods very much remain the same as hardware price increases come through.

Daniel Roberts

Maybe just to add to that as well. If we kind of fast forward and think about our strategy and capital allocation, that's part of the reason we deliberately haven't acquired 100% of the GPUs for next year. We see there's right way risk. So if you scenario map where we're at in the market at the moment, if financial markets remain weak in terms of sentiment, then -- and if construction and execution continues to be challenging across the industry, then it may well be that GPUs will fall in price, just to reflect the lack of [ power points ], purely and simply.

And then we think it's right way risk because if markets recover, they start charging. Worst case, it's correlated and maybe we've got to pay a slight premium for the GPUs. Maybe we've got to wait 2 or 3 months. But it comes back to the central thematic around it's time to compute, it's time to those data, to those [ power points ]. So prioritizing the physical infrastructure, commencing construction across all these sites around the world and preserving that optionality. If the markets are good, we'll be able to afford the GPUs, we'll be able to afford to wait an extra couple of months on the supply. But if markets remain muted, we're going to be very pleased that we haven't overcommitted on the GPU CapEx, and we're keeping the platform growth ticking along in the background.

Operator

Thank you. Ladies and gentlemen, I'm showing no further questions in the queue. I would now like to turn the call back over to Dan Roberts, Co-Founder and Co-CEO, for closing remarks.

Daniel Roberts

Thanks, everyone. To close, so Horizon 1 has been handed over to Microsoft with NVIDIA Exemplar status, 14 months from civils against an 18-month industry benchmark. And Horizons 2 to 4 are now nearing completion. The more than $4 billion of contracted IRR -- ARR, I should say -- is funded and now targeted for January. Perhaps it extends to the end of February.

This quarter, we're very proud AI cloud revenue almost tripled and our operating ARR reached $1 billion. Our backlog is over $20 billion, with 2/3 of it investment-grade counterparties. Demand, as we've mentioned many times on this call, has frankly never been stronger. In the last few days, we've contracted with a leading frontier AI lab to expand their existing capacity with us. We signed a new 5-year contract at about $22 million a megawatt, with that customer prepaying 90% of the GPU CapEx. We signed a short-term contract at $40 million a megawatt. And importantly, older generation H100s, we just recontracted at double the original contract price.

We own the land, the power, the data centers under all of it, which is what lets us set the pace. Thank you all for your time. The next milestone for us is the 0.3 gigawatts targeted for January. Thank you.

Operator

That concludes today's conference call. Thank you for your participation. You may now disconnect.