Dear Investor
The ELMRI ANZ Conviction Fund returned -2.6% in July, while the ELM Responsible Investments Global Fund returned -9.1%.
Global markets pulled back through July, as a crowded AI hardware and semiconductor trade unwound sharply after an extended period of outperformance. Hedge fund positioning in the sector had reached record levels heading into the month, and the resulting de-risking was compounded by the forced, margin driven liquidation of a large levered fund and by structural leverage in Korean and US semiconductor linked ETFs, amplifying moves well beyond what fundamentals alone would suggest. Sentiment was further weighed down by the record Shanghai debut of Chinese memory manufacturer CXMT, which reignited concerns around long term competition and oversupply risk in the memory market. Pleasingly, late month price action showed early signs of investors rotating back into previously sold off SaaS names, a tentative but welcome shift in risk appetite.
The Australian market saw a more moderate pullback over the month, with the ANZ Fund's detractors concentrated in healthcare technology and engineering services companies exposed to the broader AI buildout. Encouragingly, the first of our engineering holdings exposed to data centre construction reported a market update in the first week of August, confirming it had exceeded profit guidance, a result that has helped drive a material bounce off late July lows across several of these names.
The Australian dollar also strengthened approximately 1.5% against the US dollar over July, which detracted modestly from the Global Fund's reported AUD return by reducing the translated value of its offshore holdings.
Semiconductor Hardware: Profit Taking, Crowded Positioning and Margin Calls
July was characterised by a sharp pullback across AI hardware and semiconductor suppliers. Having served as the primary performance drivers for global equity benchmarks over recent quarters, these names experienced a pronounced momentum reversal as several compounding headwinds converged:
A crowded trade unwinds: Hedge fund exposure to semiconductors had climbed to nearly 20% of gross book on prime brokerage data, the highest level on record and more than double where it stood at the start of the year. Goldman Sachs' Prime Insights and Analytics team characterised the subsequent selling as de-risking and profit taking after a crowded trade that had worked, rather than a fundamental reassessment.
A forced levered fund unwind: A primary catalyst behind the acceleration in selling pressure in late July was technical, price insensitive liquidation rather than a breakdown in fundamentals. Reports of the forced unwind and margin call liquidation of a heavily levered AI focused hedge fund, compelled to offload a multibillion dollar public equity book, added indiscriminate selling across concentrated AI hardware and infrastructure holdings.
Leveraged ETFs: Structural leverage compounded the moves in both directions. In South Korea, retail flows into single stock leveraged ETFs on Samsung Electronics and SK Hynix amplified declines via rebalancing and margin liquidations, contributing to a near 9% single day fall in the Kospi on 13 July, while broad based 3x leveraged semiconductor ETFs magnified sector wide swings in US markets.
Emerging Chinese domestic competition: Market sentiment was further weighed down by growing concerns that Chinese semiconductor and AI competitors are gaining ground. CXMT, China's leading memory manufacturer, IPO'd on the Shanghai STAR Market on 27 July, closing up 466% on debut to become the most valuable company listed on mainland Chinese exchanges, adding to negative sentiment around long term market share and oversupply risk in the memory sector.
While short term sentiment around hardware has cooled, we view this as a natural market digestion phase within a multi year technological buildout, rather than a structural end to the cycle. Despite this near term noise and levered forced selling, operational reality on the ground remains robust. We continue to focus on the fundamental datapoints: hyperscaler capital expenditure guidance, which continues to point to sustained and, in several cases, upgraded AI infrastructure spend; forward order books and contract structures from memory suppliers, which point to demand visibility extending well beyond the typical cycle; and the structural barriers to entry that remain particularly strong across our AI hardware and semiconductor positions.
Memory: CXMT Competitive Threat
It would be inappropriate to dismiss CXMT as a competitive threat given its rapid capacity buildout and government backing as a strategic pillar of China's domestic semiconductor and AI infrastructure plans. However, we would be cautious about attributing a structural competitive advantage to the company at this stage.
CXMT remains one to two process generations behind leading global peers, relies on DUV multi patterning rather than EUV, and therefore requires materially more wafer starts per bit produced. This cost per bit disadvantage is currently masked by elevated DRAM pricing, but would become a determining margin factor once the cycle turns. Independent analyses also point to an inconsistent performance record in commodity DRAM relative to Samsung, SK Hynix and Micron, alongside a multi year gap in high bandwidth memory, where CXMT is only just establishing an initial production footprint. Export controls on advanced machinery add a further barrier. If we were analysing CXMT as a standalone investment, we would view its edge as stemming from policy support and captive domestic demand rather than superior efficiency or product quality. We continue to monitor developments given our exposure to the sector.
SaaS and Platform Software: Early Signs of Rebound as Disruption Thesis Fails to Materialise
Conversely, July provided the first tangible evidence of a fundamental rebound across high quality Software as a Service (SaaS) and enterprise platform equities. For much of the past year, software incumbents were broadly sold off under the blanket narrative that autonomous AI agents would shrink seat counts, compress pricing power and disintermediate core platforms. However, early Q2 earnings results are beginning to disprove this generic AI displacement thesis, particularly for high quality software businesses possessing the structural traits we have consistently emphasised: proprietary data, system of record status, workflow orchestration, distribution strength and pricing power.
Early earnings data disproves this displacement theory across three distinct pillars:
Accelerating agentic volume: Proof that enterprise clients are scaling actual AI deployments rather than testing prototypes. ServiceNow led this trend with a ninefold increase in active agentic workflows, while TechnologyOne has successfully scaled its agentic AI platform (Plus) across its global customer base, supporting the company's reaffirmed FY26 guidance of 18 to 20% profit before tax growth on the back of rapid public sector workflow automation.
Pricing power: Validation that incumbents are leveraging proprietary data to command contract premiums for tangible operational ROI. ServiceNow's AI native tiers are capturing 20 to 30% price uplifts over legacy packages, supported by a hybrid token based consumption model.
Material AI contract expansion: Financial evidence that AI is expanding total backlogs and revenue, not cannibalising them. ServiceNow crossed the $1 billion AI ACV milestone, with multi product AI deals growing 5.5x and the firm upgrading its end of year ACV target by 50% to $1.5 billion.
Taken together, these datapoints suggest that for platforms with proprietary data and system of record and workflow orchestration roles, AI is beginning to expand the opportunity set via agentic work units, diagnostic AI workflows and higher consumption, with early signs of AI related ACV and automation contributing to underlying revenue growth. The feared AI displacement of core software platforms has yet to materially manifest at scale, including within verticals previously thought to be highly exposed to disruption, such as professional data providers to lawyers and accountants like RELX, FactSet and Thomson Reuters. Rather than being disintermediated, these incumbents are experiencing modest revenue acceleration by embedding generative AI directly into core workflows, effectively transforming static databases into interactive, agentic software suites.
This highlights that the market is still in the very early stages of enterprises deploying inference heavy applications into production. As Amazon CEO Andy Jassy noted on the Q2 earnings call in late July, enterprises remain in the early stages of running inference at scale in production. We believe that as open models proliferate and the cost of compute falls materially over time, these dynamics will act as a catalyst for broader enterprise adoption in the medium term, with high quality SaaS and platform businesses well placed to benefit, albeit with a gradual rather than instantaneous revenue impact.
Open Weight Proliferation and the Declining Cost of Work
A defining structural shift taking place across the technology landscape in 2026 is the rapid proliferation and sophistication of open source and open weight large language models. The open source ecosystem has achieved near parity with proprietary frontier models for the vast majority of enterprise operational tasks, fundamentally resetting the unit economics of cognitive compute.
Over the past three years, the cost of processing a million tokens of intelligence has fallen by over 95%, transforming raw model intelligence from a scarce, high cost commodity into a ubiquitous utility.
Who stands to benefit?
When raw intelligence trends toward marginal cost, economic value migrates away from pure model providers and concentrates in the following domains:
Systems of record and workflow orchestrators. A language model cannot complete useful work without domain data, permission governance and an established workflow. Platforms that own the system of record can route routine tasks to cheap open weight models while reserving frontier reasoning for edge cases. This is a direct reduction in cost of goods sold at the point where AI features are being embedded, and it supports gross margin while deepening customer dependency. This is the mechanism we watch across our software holdings, including WiseTech, ServiceNow, TechnologyOne and Xero, and it is the reason we assess these businesses on revenue capture per workflow rather than on a simplistic view that AI erodes seat based licensing.
Cloud infrastructure. Self hosted and privately hosted inference still runs somewhere, generally on public cloud or hybrid capacity, and it runs continuously rather than in bursts. The hyperscalers capture recurring compute revenue regardless of whether the model weights are open or closed. The relevant test for us is return on AI capital expenditure and utilisation, not headline spend.
Compute and memory. Inference is now the larger share of AI spend and continues to grow as a proportion of the total. Cheaper tokens expand the number of workloads that clear the economic hurdle, which raises aggregate inference volume even as cost per task falls. The beneficiaries are the vendors improving tokens per watt and tokens per dollar at a system level, which keeps our focus on accelerator and high bandwidth memory exposure rather than on token volumes alone.
The end enterprise. Workflows across supply chain, customer operations and compliance that were cost prohibitive eighteen months ago now clear the hurdle rate. This is where the productivity dividend is actually created, and part of it is available for capture by the platforms that sit inside those workflows.
Model commoditisation is a catalyst, not a risk, but the transmission is slower and less direct than the headline price declines suggest. The value accrues to the layer that owns the workflow, not to the layer that owns the weights.
Company Spotlight: ServiceNow (NYSE: NOW)
ServiceNow provides a clear, real time example of how incumbent enterprise platforms are turning generative and agentic AI into compounding financial execution, providing direct evidence against the AI displacement thesis.
Key drivers behind the outperformance
AI milestone reached ahead of schedule: ServiceNow crossed $1.0 billion in AI Annual Contract Value (ACV), with net new AI ACV accelerating over 40% quarter on quarter. Driven by rapid enterprise adoption of its AI Control Tower, management confirmed the company is on track to reach $1.5 billion in AI ACV by year end.
Agentic AI and federal momentum: Agentic AI deployments expanded ninefold over a nine month window. High value enterprise adoption was reinforced by 123 large deals exceeding $1m in net new ACV (~40% year on year growth), supported by accelerated demand across the US Federal sector.
Addressing per seat disruption fears: Investors previously feared autonomous AI agents would shrink seat counts and undermine traditional SaaS pricing models. ServiceNow disclosed that 50% of its net new business is now non seat based, demonstrating that consumption and outcome linked pricing models allow platforms to capture greater value as automation increases.
Full year guidance upgrade: On the back of net new deal momentum and expanding remaining performance obligations (RPO of $29.0bn), management upgraded full year 2026 subscription revenue guidance to $15.76bn to $15.78bn.
ServiceNow illustrates that platform compounders with high switching costs, deep workflow integration and adaptable commercial models remain well positioned to compound shareholder value through this technological cycle.
Company Spotlight: Southern Cross Electrical Engineering (ASX: SXE)
SCEE's $165 million capital raise announced in June was accompanied by an upgrade to FY26 EBITDA guidance and the initiation of FY27 EBITDA guidance at at least $100m, a 33% step up. In our view the raise is best read as a signal of the depth of the data centre, renewables and electrification opportunity in front of SCEE, and the reason the balance sheet is being resized to meet it.
Insights from the update
Data centres are moving from pilot to material contributor. SCEE stated FY27 data centre revenue is expected to be approximately 3.0x FY26, is currently tendering more than $1bn of data centre work, and secured NEXTDC S4 early works via Multiplex and an LV switchroom skid contract for a major operator.
Renewables and battery storage are a distinct second growth leg. SCEE recently completed the $250m Collie battery storage project and won the Steel River East BESS award, with AEMO's ISP forecasting utility scale storage growing from 3 GW in 2024/25 to 27 GW by 2029/30.
Broad based electrification underpins the core order book. Higher electrical intensity across buildings, Rio Tinto Pilbara sustaining capital under the new three year Master Construction Agreement, and infrastructure work keep the order book diversified while data centres scale.
The guidance upgrade is grounded in delivered work. FY26 EBITDA guidance was lifted to at least $75m from at least $72m, and the FY27 step to at least $100m is supported by $150m of works awards announced with the raise and a disclosed pipeline of near term awards.
The raise is a growth raise, not balance sheet repair. Pro forma liquidity moves to approximately $308.8m, bank guarantee and surety bond capacity to $200m from $150m, and manufacturing floorspace more than doubles to over 17,000m2.
Why we continue to hold
SCEE combines direct exposure to the buildout of AI infrastructure with disciplined delivery across renewables, resources and building electrification, and management has consistently converted that exposure into earnings ahead of expectations. The read through extends to our other holdings in the engineering space, including SKS Technologies, which also delivered guidance beat in a recent trading update. Alongside IPD Group, Maas Group and Wagners, these holdings collectively span the multi year buildouts in critical infrastructure, from AI ready data centres, to the utility scale renewables, battery and transmission capacity flagged in AEMO's Integrated System Plan, to the deeper electrification of buildings and resources. All five continue to earn their place in the portfolio.
Conclusion
Near term market drawdowns across our core holdings reflect temporary shifts in style and market narrative rather than structural impairments to earnings power. While market sentiment frequently oscillates between enthusiasm for hardware enablers and scepticism toward software platforms, our process remains anchored in underlying cash generation, pricing power and durable competitive moats.
We remain disciplined, grounded and confident in the multi year compounding potential of the portfolios.
Thank you for your continued partnership and trust in ELM Responsible Investments.
Kind regards,
Jai Mirchandani
Founder, CIO and Portfolio Manager
ELM Responsible Investments
This note has been prepared by ELM Responsible Investments (‘ELMRI’) ABN 70 607 177 711 AFSL 520428, for Australian wholesale clients for the purposes of section 761G of the Corporations Act 2001 (Cth).
The information is not intended for general distribution or publication and must be retained in a confidential manner. Information contained herein consists of confidential proprietary information constituting the sole property of ELMRI and its investment activities; its use is restricted accordingly.
This note is for general informational purposes only and does not purport to be comprehensive or to give advice. The views expressed are the views of the writer at the time of preparation and presenting and all forecasts, assumptions, opinions, data and other information are not warranted as to accuracy or completeness and are subject to change without notice. This is not an offer document and does not constitute an offer or invitation of investment recommendation to distribute or purchase securities, shares, units or other interests to enter into an investment agreement. No person should rely on the content and/or act on the basis of any material contained in this note. Any potential investor should consider their own circumstances and seek professional advice.
ELMRI funds, its directors, employees, representatives and associates may have an interest in the named securities.
Past performance is for illustrative purposes only and is not indicative of future performance.