Thirty months into the generative artificial intelligence awakening, the jury is still out on how much enterprises are benefiting from investments in AI.

Though the vast majority of customers continue to spend on AI, reported returns are no greater than, and frankly lag those, typically associated with historical information technology initiatives such as enterprise resource planning, data warehousing and cloud computing. Let’s face it, the multihundred-billion-dollar annual capital expenditure outlay from hyperscalers and sovereign nations is fueling the euphoria and essentially supporting blind faith in the AI movement.

But ground truth returns from enterprise AI adoption remain opaque. This combined with geopolitical unrest, fluctuating public policy and GDP estimates in the low single digits, have buyers tempering expectations for tech spending relative to January of this year.

In this Breaking Analysis, we review the latest data on enterprise AI adoption in the context of overall IT spending. We’ll also drill into some of the dynamics of AI usage and give you a small glimpse of some leading-edge AI product research conducted by Enterprise Technology Research.

Internal AI app/dev is grabbing budget

We believe the latest cut of ETR’s AI Product Series, shown below, confirms a decisive budget shift toward building in-house AI applications. The survey, fielded to 600 IT decision makers who are actively developing and maintaining proprietary AI applications, finds that 83% expect to raise spend in 2025, with most projecting double-digit increases. Only 5% foresee any pullback. ETR notes that a companion January survey of broader “buy-and-build” AI programs delivered a similarly bullish read (75% net-positive), underscoring that whether enterprises are purchasing software as a service or crafting code themselves, AI app/dev is a budget priority.

Key findings

• Double-digit spending commitment: Our research indicates that the majority of respondents are materially expanding budgets to accelerate time-to-value for custom AI workloads.
• Consistency across buyer types: A prior, all-inclusive AI study showed nearly identical enthusiasm, signaling that demand is robust for both commercial SaaS and bespoke development paths.
• Early-cycle headroom: We believe these spending intentions represent the front end of a multi-year investment wave, given that most enterprises are still in proof-of-concept or early production stages.
• Low downside risk: With just 5 percent citing budget cuts, the data suggest AI application development is somewhat insulated from broader IT spending scrutiny. We recognize that can change with sentiment shifts.
• Cadence of validation: ETR’s regular pulse will allow us to track inflections quickly; our expectation is that the next read-out will sustain or even exceed today’s optimism; with the caveat that if ROI remains opaque (which we’ll discuss below) and public policy uncertainty remains, the sanguine sentiment could flip.

Nonetheless, in our opinion, the chart’s right-skewed distribution of responses illustrates a market that is still in acceleration mode and nowhere close to saturation. The implication is vendors supplying tooling, runtime platforms, and talent to in-house AI teams will likely see sustained tailwinds through 2026 and beyond.

Regulatory crosswinds haven’t broken AI’s momentum

The chart below paints a picture suggesting geopolitical tension and shifting policy frameworks are not derailing enterprise AI agendas. In ETR’s April 2025 macro survey of 507 leaders, half report they are either holding steady or accelerating AI programs in spite of tightening policies. Only 10% are tapping the brakes.

The horizontal bars above depict three cohorts: 1) Enterprises accelerating AI adoption; 2) Enterprises staying the course; and 3) Those slowing adoption. The middle band, labeled “proceeding cautiously,” is the largest at roughly 40%, while the deceleration slice is under 10%. The leftmost bar representing acceleration widens materially when the data is filtered for IT-telco respondents, jumping from 24% to 40%, with a corresponding lift in the “continuing” category to 26%. In our view, the data reinforces the idea that digitally intensive sectors are doubling down, not hedging.

Key findings

• Resilient demand signals: The data indicates that regulatory uncertainty is viewed more as a governance hurdle than a spending deterrent.
• Sector bifurcation: IT-telco firms display nearly 70% positive momentum, suggesting competitive pressure is crowding out caution in tech-centric verticals.
• Healthy skepticism, not stasis: The 40% “proceed with caution” cohort reflects disciplined budget stewardship rather than retreat. We believe these organizations are actively piloting use cases while monitoring evolving compliance frameworks. We see this cohort however as similar to the swing vote in elections – if sentiment could change so could their spending patterns.
• Limited downside without a catalyst: With deceleration stuck at 10%, the probability of a broad AI investment freeze appears lower. Again, we’ve seen how changes in policy or market conditions can negatively affect budget allocations.
• Optionality remains: Undecided buyers can still tip positive; their budgets are allocated, but release is gated to clearer ROI metrics.

In our opinion, the chart’s asymmetry – that is, mostly “green” with a narrow “red band” cutting, signals that AI has earned prominent status in 2025 budgets. The prudent middle will either migrate toward acceleration as regulatory guardrails solidify; or stand pat if uncertainty persists. We caution that spending is fickle and often ebbs and flows with economic sentiment.

ROI lags the hype, revealing immaturity of the AI cycle

The chart below injects a cautionary note into the otherwise bullish narrative. Thirty months into the generative-AI wave, 27% of respondents acknowledge they have yet to see any tangible return on their AI spend. Fewer than 10% report an ROI north of 25%, the hurdle rate many CFOs use to green-light higher risk IT projects. Even more revealing is that more than one-quarter admit they are not tracking ROI at all.

The bar segments above show several ROI bands ranging from Not tracking, Negative/Zero, 0–10 %, 10–25 %, and three bands “>25 %.” The largest cohort are either not measuring or have yet to cross break-even. The data underscores a skew toward unproven or unmeasured outcomes.

Key findings

• Proof-of-concept phase: Despite vendor narratives to the contrary, our research indicates enterprises are still in experimentation mode, allocating funds to build, test, and learn rather than to harvest immediate returns.
• Budget reallocation, not expansion: Qualitative feedback and other quantitative survey data suggests AI spending often cannibalizes line-of-business, sales, and marketing budgets, delaying formal ROI scrutiny.
• Deferred accountability: The 27% “not tracking” cohort highlights a mindset of “build now, justify later,” especially among firms that maintain their own codebases.
• ROI gap vs. hurdle rates: With sub-10% clearing a 25% ROI bar, we believe many projects could soon face CFO pressure to articulate value beyond efficiency anecdotes.
• Talent and opex offsets: In addition, we note that slower head-count growth is viewed internally as an implicit payback mechanism, perhaps further postponing rigorous financial scrutiny.

Once again we see an asymmetry in the data – that is, a large cohort at “no return” while tapering sharply at “high return.” In our view this signals that AI remains an early-cycle investment story. The “proceed with caution” middle we highlighted earlier could flip positive once tangible benefits surface; conversely, it could sour if ROI continues to lag expectations or economic sentiment shifts. Either way, the next two survey waves will be pivotal in determining whether optimism converts into measurable economic value.

AI is not immune to cuts: CFOs will trim experimental spend first

We believe the chart below offers an important reality check. When 219 C-suite executives were asked where they would pull dollars if macro conditions deteriorate further, AI initiatives rank as the second-most “very vulnerable” category — trailing only outsourced IT services and sitting behind security cloud, software as a service and core infrastructure.

The horizontal bars above rank seven IT domains by share of executives evaluating the vulnerability of each segment to incremental cuts. Outsourced IT services tops the list, followed by AI projects. Security anchors the group with the smallest exposure, while cloud computing, SaaS and foundational infrastructure occupy the middle ground, indicating to us that they are somewhat more insulated.

Key findings

In our opinion, the data reminds investors and practitioners that AI hype does not equal budget immunity. While earlier data documented robust intent to invest, the C-suite view presented above highlights a Caveat. Specifically, if macro headwinds intensify, experimental AI funding could move to the front of the line for cutbacks.

IT budget growth retreats as macro sentiment shifts

The data below captures one of the sharpest one quarter resets in enterprise tech spending since the initial COVID shock. ETR’s April macro pulse shows 2025 IT budget growth expectations plunging from 5.3% in January to 3.4%, a 190-basis-point drop which ETR characterizes as “the biggest drop we’ve ever recorded outside a crisis.”

The line graph above tracks sequential quarterly revisions to aggregate IT budget growth, with call-outs for key segments. All three curves slope steeply downward between Q1 and Q2. The mainline average falls to 3.4 percent, while the Fortune 500 and Global 2000 are even less optimistic, looking for 2.4% and 2.2%, respectively. One footnote is the survey closed one day after reciprocal tariffs were announced, meaning the data miss the subsequent equity market sell-off that likely would have driven estimates even lower. The sample also was taken before the v-shaped rebound that ensued when tariffs were put on temporary hold.

Key findings

In our opinion, the data’s pronounced downward inflection delivers two messages. First, the post-COVID era of “high-single-digit baseline” IT growth is over for now; budgets have snapped back to pre-pandemic norms. Second, the gravity of macro sentiment is palpable. If GDP stays sub-1%, we expect CFOs will continue to recalibrate spend to 2% to 3% territory, forcing AI champions to prove value faster or risk funding erosion.

Policy fog causes the C-suite to tap the brakes on net-new IT projects

The data below is the clearest signal yet that political and trade turbulence is starting to seep into executive decision making. In a flash response cut of 219 C-level executives, 71% acknowledge some form of pullback, whether delaying, reducing scope, pausing or outright canceling AI initiatives, while only 27% say uncertainty pushed them to accelerate or green-light new projects.

Key findings

In our opinion, the chart’s lopsided distribution toward delay and down-scoping underscores an emerging tension between boardroom optimism about AI’s transformative potential versus CFO caution in an unsettled macro climate. The next survey wave will reveal whether today’s pause becomes tomorrow’s cancel, or merely a coiled spring set to release once policy headwinds subside.

Enterprise LLM app/dev adoption shows GPT alone at the top

The inaugural ETR snapshot of which foundation model families enterprises plan to incorporate into their own code bases provides an early but telling signal — that is, OpenAI’s GPT commands unprecedented mindshare, with roughly 62% of builders already using or planning to use the model. No rival cracks the 35% line, underscoring GPT’s first-mover momentum and Microsoft’s distribution advantage through Azure OpenAI Service.

The horizontal bar chart shows more than a dozen model families from highest to lowest “currently using/plan to use” share. GPT’s bar stretches almost two-thirds of the way across the graphic, colored solid green for active use and patterned green for near-term intent. Meanwhile the next cluster (Gemini, Anthropic Claude, Llama and so on) hovers in the high 20% to low 30% range. A tapering long tail follows with Databricks’ DBRX, Amazon’s Nova, Snowflake Arctic, DeepSeek, Grok, IBM Granite, and Cohere each registering lower but notable interest.

Key findings

In our opinion, the chart’s steep drop-off from GPT to “everyone else” confirms that distribution and developer experience matter as much as raw model performance in the enterprise race. Yet the presence of nine alternative families, many embedded directly inside data platforms, signals a coming phase of fragmentation and price competition once baseline accuracy converges.

Microsoft extends its lead in AI development platforms, but optionality remains critical

We believe the data below illustrates a familiar pattern — that is, the farther enterprises move up the AI stack, the more dominant Microsoft becomes. When respondents were asked which development platforms they are using or plan to use for building AI applications, three Azure offerings occupy the top three positions.

Following Microsoft’s dominant adoption, the chart below shows Amazon’s Bedrock (mid 30%) and SageMaker (low 30%) form the second tier, followed by Google Vertex AI in the 30% range. A mixed long tail rounds out the chart, including Hugging Face, LangChain, Databricks Mosaic AI, Snowflake ML/Cortex, IBM watsonx, Oracle AI, Cloudflare Workers AI, Red Hat OpenShift AI and H2O.ai.

Key findings

• Azure as de-facto starting point: The data indicates that Microsoft’s simplified handoff from GPT to ML tooling provides a facile onramp that few rivals can currently match, at least in terms of account capture.
• Public-cloud friction is real: CIOs at heavily regulated firms told ETR they cannot put sensitive data in public-cloud AI services, underscoring why on-platform options register meaningful interest.
• AWS shows a two-track strategy: Bedrock’s model garden attracts greenfield generative-AI projects, while SageMaker continues to serve classical ML workloads. Together and combined with likely higher ASPs protect against Azure’s account penetration.
• Google Vertex gains on developer experience: Despite lagging GPT in raw hype, Vertex’s integrated tooling and first-party Gemini access place it firmly in the mix, especially for cloud-native shops.
• Long-tail resilience through proximity to data: We believe offerings embedded directly in data platforms like Snowflake Cortex and Databricks Mosaic, punch above their weight because they minimize data movement and governance overhead.
• Compliance drives optionality: The sheer breadth of vendor logos in the chart reflects enterprise demand for deployment models that span public cloud, dedicated region, on-prem and hybrid Kubernetes estates.

In our opinion, the data underscores a two-fold conclusion: 1) Microsoft’s end-to-end positioning from model to tooling is paying off; yet 2) The market is far from winner-take-all because data-sovereignty, latency and existing platform commitments compel a diversified toolkit. Vendors that can meet customers where their data already lives — while abstracting away model orchestration complexity will gain outsized leverage as the AI stack evolves.

Vector search has shifted from hot topic to table stakes, but the field remains crowded

We believe the data below demonstrates how rapidly vector search has moved from headline innovation to checkbox feature – and how that shift has splintered adoption across a surprisingly wide set of vendors. When asked which vector-search engines they are currently using or plan to use, respondents placed Microsoft’s Azure AI Search narrowly in front, followed by MongoDB’s native Atlas Vector Search and an almost dead-heat between Elasticsearch and Redis. A second cluster comprising Google’s Vertex AI, Amazon Kendra, Databricks’ Vector Store, Snowflake Arctic Vector and Oracle’s OCI AI Search, occupies the mid-teens, followed by IBM watsonx.data (IBM uses Milvus) and specialist standalone offerings such as Pinecone, Milvus and Cloudflare Workers AI round out a long tail of single-digit intent.

Key findings

In our opinion, the data confirms that vector search has crossed the chasm into “must-have” territory, yet remains far from winner-take-all. Enterprises still evaluate options by asking, Which engine fits my existing data topology and latency SLAs? rather than defaulting to a single cloud offering. We expect consolidation to follow two paths: 1) A deeper embedding of vector indexes inside mainstream databases; and 2) Specialist engines targeting high-throughput, cross-region retrieval-augmented-generation or RAG at lower cost.

Skills, data quality and legacy integration form constraints on DIY AI

What the chart shows

A vertical bar chart ranks these hurdles left to right. The first three bars — skills, data quality, legacy integration — tower over the rest, followed by a mid-pack of compliance, budget and scaling, then a taper toward soft factors such as culture.

The data underscores why large enterprises find do-it-yourself AI far harder than the marketing demos imply. In a 600-respondent data cut focused on organizations building their own apps, several obstacles jump out as shown below, led by skills, data quality and integration challenges.

Key findings

• Build-over-buy skews big-company: Our research indicates Fortune 500 firms prefer DIY stacks for customization and IP control, but that preference amplifies the constraints — talent, data prep and legacy plumbing — they rate most painful.
• Data remains the gating asset: Without reliable lineage, masking, and vector-friendly schemas, fine-tuned models underdeliver. This in our view validates the idea of “bring AI to the data, not the data to AI” narrative we outlined in prior research (No. 4 from our 2025 predictions post).
• Compliance casts a long shadow: Security teams still view public-cloud inference as riskier than on-prem sandboxing, pushing organizations toward private deployments even as GPU scarcity persists.
• Cost is not the primary barrier as of yet: Only 30% cite budget as a top-three concern, implying CFO scrutiny will intensify once initial pilots transition to production-scale clusters.
• Talent wars could decide platform winners: Vendors that bundle turnkey MLOps, lineage and guardrails will siphon demand from DIY pipelines as hiring constraints eat into talent pools.

In our opinion, the data underscores a clear truth – that is, AI’s limiting factors are less about technological breakthroughs and more about enterprise plumbing. Solving for data hygiene, skills acquisition and legacy connectivity will matter more to adoption velocity than squeezing a few extra tokens per dollar out of the next model release.

Bring AI to the data: Deployment patterns split between cloud convenience and on-prem control

We believe the data below validates a core theme running through this research note – that is, that most enterprises are bifurcating their AI estates, keeping sensitive workloads close to the data while offloading everything else to the public cloud.

Key findings

• Cloud dominates — but not universally. Our research indicates that more than half of AI workloads spin up in hyperscaler regions thanks to turnkey tooling, elastic GPUs and mature MLOps services.
• On-prem remains mission-critical. Roughly one-third of respondents split between private-cloud and legacy on-prem deployments, driven by data-gravity, latency and regulatory mandates. This is particularly true in financial services, insurance and healthcare.
• Private cloud is different than legacy on-prem. Private-cloud environments deliver cloud-like APIs and self-service portals behind the firewall. Our interpretation is that “on-prem” in this survey refers to classic siloed estates lacking unified storage/compute/network control planes. Scale, not geography, is the distinguishing factor.
• Dell and HPE lead the on-prem race. When respondents were asked which vendors they would choose for in-house AI stacks, Dell Technologies and Hewlett Packard Enterprise topped the list, confirming their early traction with “AI factory” bundles.
• GPU-aaS is a sleeper segment. Only about 2% of workloads land on GPU utility clouds today, yet interest is climbing; Lambda Labs and CoreWeave out-poll several Tier-2 hyperscalers for burst capacity.
• Edge is nascent but inevitable. Retail point-of-sale, robotic inspection, and autonomous-vehicle inference account for a small slice today, but we believe edge share will double once AI models shrink and 5G private wireless densifies. Notably, cloud players (for example, Amazon Lambda, Azure Functions) are prominently cited by respondents.

In our opinion, the chart crystallizes the emerging “bring AI to the data” narrative. Public-cloud agility remains unmatched for rapid prototyping, yet sovereignty-minded sectors refuse to copy/move exabytes of sensitive data into someone else’s region. Vendors that can deliver a cloud-like experiences everywhere (supercloud) – for example, public region, colocation cage, data-center row or factory floor – will capture the next tranche of enterprise spend. The race is on to make on-prem AI as turnkey as SaaS while preserving the governance guarantees boards now demand.

The AI data center super-cycle is real: Cloud leads, but on-prem will top $100B by 2030

We believe the data below confirms that generative AI has ignited the largest single-purpose infrastructure boom since the client-server era. The stacked area forecast — split into dark-blue “AI in the cloud” and light-blue “AI on-prem in the enterprise” — shows aggregate AI-only data-center spend (compute, storage, networking, power and cooling) vaulting from roughly $43 billion in 2023 to $180 billion in 2024, surpassing $250 billion in 2025 and kicking off what we’ve called a data center super-cycle.

Our model projects a 24% 10-year CAGR for AI data center buildouts through 2034, but an even faster 40% CAGR for enterprise on-prem deployments. By 2030, we estimate on-prem AI data center infrastructure will eclipse $100 billion, overtaking today’s entire footprint of general-purpose data-center gear and carving out a market that rivals early hyperscale spend profiles. Importantly, this view excludes SaaS consumption, underscoring just how massive the physical-infrastructure TAM is before a single subscription dollar is counted.

Key findings

In our view, the chart’s widening light-blue band crystallizes the “bring AI to the data” thesis. On-prem estates may have ceded generic workloads to hyperscalers, but they will reclaim budget as organizations embed AI next to proprietary datasets that can’t, or won’t, move. Vendors that deliver turnkey AI factories with cloud-like control planes, rich data-stack partnerships, and ecosystem breadth will be positioned to capture the next $100 billion wave.

How do you see it? Where are you putting your AI bets? Let us know and thanks for reading.

Disclaimer: All statements made regarding companies or securities are strictly beliefs, points of view and opinions held by SiliconANGLE Media, Enterprise Technology Research, other guests on theCUBE and guest writers. Such statements are not recommendations by these individuals to buy, sell or hold any security. The content presented does not constitute investment advice and should not be used as the basis for any investment decision. You and only you are responsible for your investment decisions.

Disclosure: Many of the companies cited in Breaking Analysis are sponsors of theCUBE and/or clients of theCUBE Research. None of these firms or other companies have any editorial control over or advanced viewing of what’s published in Breaking Analysis.

Image: theCUBE Research/DALL-E