“Report by BN HQ for comms, you will be DZSO for RRC’s night HALO.”
“Q2 issues included negative hyperscaler FCF, HBM supply constraints throttling inference scalability and surging GPU depreciation headwinds.”
What do the above two sentences have in common? They both appear to be written in English, but you really have to speak another language to understand them. They both have negative tone. In the case of the first sentence, I used to get quite depressed when I got tagged to be Drop Zone Safety Officer for a night jump – which meant my afternoon and night were blown and I’d be spending the next day trying to stay awake. Definitely negative tone! Unfortunately, I don’t speak the language of the second sentence, but it references concerning AI themes that have emerged during the Q2 earnings season. With the changes that are impacting the AI trade, I thought it was important to communicate that we are entering a new phase in the trade, but I realized that I don’t fully speak the language or understand the interrelationships between the major AI companies. This commentary attempts to clarify some of the jargon commonly used in the AI discussion and then address some of the significant changes in the AI landscape that affect us as investors. Finally, I apologize for the use of terms like “parallel,” “series,” and “NAND;” I know that there are some of you who, like me, are still traumatized from taking EE362 (Electrical Engineering: Digital Logic).
Despite focusing on some of the emerging concerns within the AI ecosystem, we remain constructive on the sector.
For the past several years, markets have generally rewarded companies for simply announcing AI infrastructure spending or Capital Expenditures (CapEx). The Q2 2026 earnings season seems to have ended that and entered a new “show me the money” phase. Investors are now focused on whether the hundreds of billions being spent on CapEx are generating real revenue. Some companies that reported increased Q2 CapEx along with negative free cash flow (FCF) were punished severely by the market. We will address this change, as well as several other themes that have emerged from the Q2 earnings season, but first let’s make sure that we understand the key terms, categories of AI companies, and their interrelationships.
Like the military, the AI industry has its own language. The below is a list of key terms that should help us understand the AI language:
|
Term |
Plain-English Definition |
|
Generative AI (GenAI) |
A type of AI that can create new content — text, images, code, or video — based on patterns learned from vast amounts of data. ChatGPT is the most famous example. |
|
Large Language Model (LLM) |
The underlying technology behind tools like ChatGPT or Google Gemini. These are AI systems trained on enormous amounts of text to understand and generate human language. |
|
AI Agent / Agentic AI |
An AI system that can take actions autonomously — browsing the web, writing code, or managing tasks — without constant human instruction. The next frontier beyond simple chatbots. |
|
GPU (Graphics Processing Unit) |
Originally designed for video games, GPUs are now the primary "engine" used to train and run AI models. They are specialized processors optimized for parallel computing – performing thousands of simpler mathematical operations simultaneously. Nvidia is the dominant supplier. |
|
CPU (Central Processing Unit)
HBM (High Bandwidth Memory) |
CPUs are general-purpose processors optimized for series or sequential computing. As AI has shifted from training to inference, CPUs have become more important with the need for CPUs and GPUs trending toward equality. Intel, AMD, Arm Holdings and Nvidia are the primary designers. A specialized, ultra-fast type of computer memory that sits alongside GPUs to feed them data quickly. SK Hynix, Samsung, and Micron are the main producers. The most critical bottleneck in AI supply chains. Dynamic Random Access Memory (DRAM) and NAND Flash Memory are also significant in the Ecosystem. |
|
Data Center |
Large facilities filled with servers, GPUs, and networking equipment that store data and run AI computations. The physical backbone of the AI economy. |
|
Hyperscaler |
The largest cloud computing companies — Amazon (AWS), Microsoft (Azure), Google (Google Cloud), and Meta — that operate massive data centers and sell computing power to businesses worldwide. |
|
Cloud Computing |
Renting computing power, storage, and software over the internet rather than owning physical hardware. AWS, Azure, and Google Cloud are the three dominant providers. |
|
Foundry / Fab |
A factory that manufactures chips. TSMC (Taiwan Semiconductor) is the world's most important foundry, making chips for Nvidia, Apple, and others. The relationship between TSMC and Nvidia is the single most consequential relationship in the AI ecosystem. Any disruption to TSMC’s Taiwan operations would halt Nvidia GPU production, with no near-term alternative. |
|
Inference |
The process of running an already-trained AI model to generate answers or outputs. As AI moves from development to deployment, inference demand is growing rapidly. |
|
Training
Tokens |
The process of teaching an AI model by feeding it vast amounts of data. Extremely compute-intensive and expensive — the primary driver of early GPU demand. A token is the basic unit of text that an AI language model reads and generates. A token can be a whole word, part of a word, punctuation, spaces, etc. AI model providers price API access in dollars per million tokens processed. |
|
AI Accelerator |
Any specialized processor designed to perform AI computations. They include GPUs, Application-Specific Integrated Circuits (ASIC) which are custom chips designed for a specific AI workload and Tensor Processing Units (TPU) which is also a custom chip designed to accelerate machine learning workloads. |
|
Optical Transceiver
EUV (Extreme Ultraviolet) Lithography
|
A component inside data centers that transmits data at high speeds using light. US vendors Lumentum, Coherent, and Applied Optoelectronics hold about 25% of the global optical transceiver market. Chinese OEMs hold over 50% of the global market. A current geopolitical flashpoint, as the US is considering banning Chinese-made versions. The most advanced chip-printing technology using short-wavelength light to etch circuit patterns onto silicon wafers. ASML is the sole manufacturer of EUV lithography machines in the world making ASML one of the most strategically critical companies in the entire semiconductor supply chain. |
The AI economy is not a single industry — it is a layered ecosystem of interdependent companies. Figure 1 is a map of the major categories and the key companies within each.
Figure 1
Source: AAFMAA Wealth Management & Trust. We created this chart with the help of Bloomberg AskB, an interesting exercise where we worked with AI itself to help map the AI ecosystem.
As you can see from Figure 1, the number of companies and interrelationships between them is overwhelming. While it is beyond the scope of this commentary to go through each category and company, the next sections simplify the AI Ecosystem into a smaller number of more manageable layers and discuss what they do and the major developments that we’ve seen so far in Q2.
Layer 1 companies embed AI capabilities into enterprise software that end-users interact with daily. They are the monetization frontier of the AI stack — the layer where the Return on Investment (ROI) question is most acute and where AI spending must translate into measurable productivity gains.
Microsoft — Copilot embedded across Microsoft 365, GitHub, and Azure; the broadest AI application surface area of any enterprise software company
Salesforce — Agentforce and Einstein AI embedded across CRM, Sales Cloud, and Service Cloud
ServiceNow — AI workflows embedded across IT service management, HR, and enterprise operations
Palantir — AIP (Artificial Intelligence Platform) deployed for commercial and government decision-making and operational AI
Layer 2 companies build and train the large language models (LLMs) that power virtually all AI applications. They are the intellectual core of the AI stack — developing foundational models accessed via Application Programming Interface (API), embedded in cloud platforms, or deployed as open-source weights. This layer is characterized by intense competition, rapid capability improvement, and significant capital consumption.
OpenAI — GPT-4o and o-series reasoning models; the dominant commercial LLM provider by revenue
Anthropic — Claude model family; leading in enterprise safety and long-context performance
Google DeepMind — Gemini model family; uniquely integrated with Google Search, Cloud, and advertising
Meta — Llama open-source model family; the dominant open-weight model ecosystem
xAI — Grok models; integrated with X (Twitter) and targeting real-time data advantages
The hyperscalers operate the massive data centers that rent AI computing power to model developers, enterprises, and governments. They are the primary distribution layer for AI infrastructure — buying GPUs by the hundreds of thousands, building data centers to house them, and renting that capacity via cloud platforms. They are simultaneously the largest customers of the semiconductor and data center layers below them and the primary revenue channel for the model layer above.
Amazon Web Services (AWS) – largest cloud provider; Tranium and Inferential custom AI chips; Bedrock model platform
Microsoft Azure — exclusive OpenAI cloud partner; Azure AI platform; Maia custom ASIC
Google Cloud — TPU custom silicon; Vertex AI platform; Gemini API distribution
Meta — building the world's largest private AI compute cluster; exploring third-party compute rental
Oracle Cloud — fastest-growing hyperscaler by backlog; preferred by sovereign AI and GPU-as-a-service customers
Layer 4 companies provide the high-speed interconnects, switches, and load balancers that move data between servers within AI clusters. As GPU clusters scale to tens of thousands of accelerators, networking becomes a critical bottleneck — the speed and latency of data movement between GPUs directly determines training throughput and inference efficiency.
Arista Networks — dominant provider of high-speed Ethernet switching for AI backend networks; scale-out and scale-across AI fabric architectures
Cisco — Silicon One switching systems and Acacia optics for hyperscaler AI infrastructure; enterprise AI networking
Nvidia (InfiniBand & Spectrum-X) — InfiniBand remains the dominant AI backend networking protocol for training; Spectrum-X Ethernet for AI inference
Layer 5 companies design and manufacture the GPUs, TPUs, and custom AI ASICs that perform model training and inference. Nvidia dominates with its GPU architecture and CUDA software ecosystem, while AMD competes with its Instinct GPU line, and hyperscalers develop proprietary ASICs to reduce Nvidia dependency.
Nvidia — H100/Blackwell/Vera Rubin GPU families; CUDA software ecosystem; NVLink, InfiniBand, and Spectrum-X networking; Vera CPU for agentic AI
AMD — Instinct MI300X/MI350/MI450 GPU accelerators; EPYC server CPUs
Intel — Xeon server CPUs; Gaudi AI accelerators; custom ASIC design; Intel Foundry 18A
Google — TPU v5/v6 custom AI accelerators; 2–3x cost advantage over Nvidia GPUs for inference
Amazon — Trainium (training) and Inferentia (inference) custom ASICs; Graviton server CPUs
Broadcom — custom ASIC design for hyperscalers (Google TPU, Meta MTIA); PAM4 Digital Signal Processors (DSP) for networking
Layer 6 companies produce the memory and storage that feed data to AI accelerators. The memory market has bifurcated into HBM — co-packaged with GPUs for ultra-high-speed data access — and conventional DRAM and NAND flash, which serve as system memory and storage for AI servers, context buffering, and KV cache management. Both segments are in acute shortage driven by AI demand.
SK Hynix — approximately 62% HBM market share; sole HBM supplier to Nvidia for multiple generations; HBM4 in mass production
Samsung — HBM4 qualification with Nvidia achieved; largest DRAM and NAND producer by volume
Micron — HBM4 in high-volume production; SOCAMM (Small Outline Compression Attached Memory Module) / LPDDR5X (Low Power Double Data Rate 5X) leadership for agentic AI server memory
Western Digital — enterprise SSD (Solid State Drive) and HDD (Hard Disk Drive) for AI data center storage
Seagate — nearline HDD for AI training data storage
Layer 7 companies provide the physical infrastructure that houses all AI compute — the facilities, power systems, cooling equipment, and servers that make the entire stack operational. Power availability, cooling capacity, and physical construction timelines now constrain GPU deployment as much as chip supply, making this layer a critical bottleneck in the AI buildout.
Equinix & Digital Realty — colocation data center REITs providing space, power, and interconnection to hyperscalers and enterprises
Vertiv — power and thermal management systems (UPS, PDUs, liquid cooling) for AI data centers
Schneider Electric — electrical distribution, power management, and cooling infrastructure
Eaton — electrical components and power management; liquid cooling systems
Quanta Services — data center construction and electrical infrastructure
Quanta Computer / Foxconn — AI server and rack-scale system manufacturing
The Q2 2026 earnings season delivered results that, on the surface, were extraordinary by any historical measure. Aggregate cloud revenue for the major US hyperscalers reached $106.3 billion, with year-over-year growth accelerating to 43%. Aggregate cloud operating profit reached $41.4 billion, a 65% year-over-year increase, with stable to higher margins overall. The Q2 2026 earnings season was the strongest in years for US companies broadly, with projected blended annual EPS growth for the S&P 500 around 40%, and technology sector earnings projected to rise 57% in 2026. At the application layer, Palantir reported Q2 revenue up 93% year-over-year, ServiceNow surpassed $1 billion in AI Annual Contract Value, and Microsoft 365 Copilot surpassed 30 million paid seats. These results represent the clearest evidence yet that enterprise AI monetization has arrived.
Yet beneath these headline numbers, a more complicated and in some respects troubling picture emerged — one that the market began to price in during the quarter, with semiconductor and AI-related stocks experiencing a sharp sell-off even as results beat consensus. Below we address the major sources of concern that we will be following to identify potential stress on the AI trade:
The most structurally significant development of the Q2 2026 earnings season was the simultaneous deterioration of free cash flow across multiple hyperscalers — a development with no historical precedent at this scale. Alphabet reported negative free cash flow of $5.9 billion in Q2 2026 driven by capital expenditures of $44.9 billion, roughly double the prior year. Management explicitly stated that FCF is expected to remain under pressure. Meta's FCF was essentially zero in Q2 2026, with $31.9 billion in cash from operations exactly offset by capital investment. FCF is expected to turn negative in Q3 2026 and thereafter, with FY26 CapEx guided to $130–145 billion — a 100% year-over-year increase. Amazon's trailing twelve-month FCF turned negative at -$7.6 billion, a significant deterioration from positive $18.2 billion a year prior, with CapEx raised by $20 billion due to memory inflation. The entire AI supply chain is priced on the assumption that hyperscaler CapEx continues to grow. With FCF turning negative, the possibility of a single major hyperscaler Capex guidance cut is higher and if it materialized could potentially propagate instantly through every layer of the ecosystem.
The AI buildout has become the dominant force in global credit markets, and the scale of issuance is beginning to strain absorption capacity in ways that could trigger a disorderly repricing. Companies involved in the AI arms race may need to raise $1.5 trillion in investment-grade bonds over the next five years, with credit markets introducing concentration risk as investors absorb an increasingly large and correlated technology capital-spending pipeline. Oracle's five-year credit default swap reached approximately 203 basis points — the highest since late 2008 — and CoreWeave had to increase the interest rate on a $2.6 billion loan tied to Anthropic contracts to approximately 9.1% with stronger lender protections. A failed bond deal, a ratings downgrade of a major hyperscaler, or a covenant breach at a GPU cloud provider such as CoreWeave could immediately widen spreads across all AI infrastructure debt and trigger equity de-rating.
As mentioned earlier, the Nvidia/TSMC relationship is the single point of failure for the global AI supply chain. Any credible threat to TSMC's operational continuity would trigger an immediate and severe sell-off. A Taiwan Strait incident would simultaneously hit every fabless chip designer — Nvidia, AMD, Broadcom, Qualcomm — as well as TSMC itself and every company whose product roadmap depends on leading-edge silicon, which is effectively the entire AI ecosystem. Escalating export controls, the potential ban on imported Chinese-made optical transceivers and any other action that puts pressure on US China relations will remain under intense scrutiny.
Power availability is now the binding constraint on AI infrastructure deployment, and a systemic power failure or a coordinated announcement of data center delays would reprice the entire infrastructure buildout timeline. Approximately 30–50% of data center capacity scheduled for completion in 2026 may be delayed to 2027–2028 due to power constraints, and power issues are expected to lead to bankruptcies in the data center space as businesses wait for adaptations to power capacity. At least 75 data center projects worth approximately $130 billion were blocked or delayed in Q1 2026 due to community backlash and AI capacity constraints. A major utility announcing it cannot serve planned data center loads in a key market such as Northern Virginia, Texas, or Arizona could immediately hit data center REITs, hyperscalers, and GPU cloud providers whose revenue timelines depend on on-schedule capacity delivery.
The DeepSeek shock of January 2026 demonstrated that a single Chinese model release can trigger an immediate, broad-based sell-off across the entire AI supply chain — from hyperscalers to GPU makers to memory producers. The mechanism is straightforward: if frontier AI capability can be achieved at a fraction of the compute cost assumed by the market, the entire CapEx thesis collapses simultaneously. A Chinese startup named Moonshot introduced a Kimi K3 model claimed to be comparable to AI offerings from OpenAI and Anthropic, contributing to a semiconductor sell-off already influenced by valuation and CapEx-payback concerns. DeepSeek's DSpark breakthrough — a technique that accelerates inference tasks — enables continued development of domestic AI accelerator chips despite US embargoes and highlights China's leadership in low-cost AI model implementation. Reports that DeepSeek is developing its own AI inference chip triggered a fresh wave of chip-sector selling, dragging Asian markets broadly lower. China's latest AI model releases suggest the country is narrowing the capability gap with leading US labs while maintaining cost advantages. A credible Chinese model at 10–20% of Western development cost could simultaneously hit Nvidia on the GPU demand thesis, hyperscalers on CapEx justification, memory producers on high bandwidth memory demand, and AI application companies on pricing power.
Regulatory risk has moved from theoretical to active during 2026, with material financial penalties already imposed and structural remedies under discussion. Google was fined approximately €9.5 billion for abuse of dominance and an additional €890 million for Digital Markets Act violations. Potential antitrust allegations concerning Alphabet's search, ad exchange, and Android business could lead to a company breakup, with AI regulation worldwide raising issues regarding content presentation and data privacy for generative AI products. Increasing regulatory scrutiny — specifically antitrust concerns — is cited as a downside risk for Microsoft. A court order requiring Google or Microsoft to divest AI assets, or a government-mandated halt to a frontier model release, could trigger immediate repricing of AI application and platform companies and raise the regulatory risk premium across the sector.
The Q2 2026 earnings season confirmed that the AI buildout is real, accelerating, and generating measurable revenue. But it also surfaced a set of structural risks that, taken together, represent a credible scenario for a significant market correction in AI-exposed equities.
The simultaneous deterioration of free cash flow across the four largest hyperscalers, the emergence of Chinese frontier models at a fraction of US development costs, the physical impossibility of delivering sufficient power to planned data centers on schedule, the unprecedented scale of AI-related debt issuance straining credit markets, and a regulatory environment that is tightening on multiple continents are present-tense developments that emerged clearly from Q2 2026 earnings disclosures.
The central question heading into the second half of 2026 is whether AI revenue growth can accelerate fast enough to justify a capital expenditure cycle that is now consuming more cash than the world's most profitable companies can generate. The answer to that question will determine whether the AI trade continues to broaden — or whether the market begins to price in a more difficult path to returns than the current consensus assumes.
Although we have focused on emerging concerns in this commentary, we have not turned negative on the AI trade or the power of AI to transform the world. We do believe, however, that market participants have become more mature in their analysis of AI investment risks, asking companies to “show me the money” and creating an investment environment that more carefully balances risk and reward.
Yours in trust,
At AAFMAA Wealth Management & Trust LLC, we are committed to serving the unique financial needs of the military community. Whether you require a complimentary portfolio review, a comprehensive financial plan, assistance with your investment strategy, or trust services, our military wealth management professionals are ready to serve you. Contact us today to set up an appointment with a Relationship Manager who can assess your financial health and customize your personalized action plan.
Founded in 2012, AAFMAA Wealth Management & Trust LLC (AWM&T) was created to meet the distinct financial needs of military families. We proudly deliver experienced, trustworthy financial planning, investment management, and trust administration services – all designed to promote lasting security and independence.
We are proud to share the mission, vision, and values of Armed Forces Mutual, our parent company. We consistently build on the Association’s rich history and tradition to provide our Members with a source of compassion, trust, and protection. At AWM&T, we are committed to serving as your trusted fiduciary, always putting your best interests first. Through Armed Forces Mutual's legacy and our financial guidance, we provide personalized wealth management solutions to military families across generations.
© 2026 AAFMAA Wealth Management & Trust LLC. Information provided by AAFMAA Wealth Management & Trust LLC is not intended to be tax or legal advice. Nothing contained in this communication should be interpreted as such. We encourage you to seek guidance from your tax or legal advisor. Past performance does not guarantee future results. Investments are not FDIC or SIPC insured, are not deposits, nor are they insured by, issued by, or guaranteed by obligations of any government agency or any bank, and they involve risk including possible loss of principal. No information provided herein is intended as personal investment advice or financial recommendation and should not be interpreted as such. The information provided reflects the general views of AAFMAA Wealth Management and Trust LLC but may not reflect client recommendations, investment strategies, or performance. Current and future financial environments may not reflect those illustrated here. Views of AAFMAA Wealth Management & Trust LLC may change based on new information or considerations.