The headline from the launch is speed — Google says 3.5 Flash runs four times faster on output tokens per second than comparable frontier models. But the more consequential claim is the agentic performance data. On Terminal-Bench 2.1, the benchmark most closely aligned with autonomous coding tasks, 3.5 Flash scores 76.2%, edging ahead of Gemini 3.1 Pro’s 70.3% and trailing only GPT-5.5 at 78.2%. On MCP Atlas, which tests multi-step tool coordination, it leads the field at 83.6%. The model that costs roughly 25% less than its predecessor is beating that predecessor where it matters most for production deployments.

“It’s clear we’re firmly in our agentic Gemini era.”
— Sundar Pichai, CEO, Google/Alphabet, Google I/O Keynote, May 19, 2026

What the Benchmarks Actually Measure

Terminal-Bench 2.1 tests a model’s ability to operate a computer through a terminal — executing multi-step tasks, using tools, and recovering from failures without human intervention. MCP Atlas evaluates deterministic tool orchestration across complex, multi-agent pipelines. These are not academic exercises. They are close proxies for the workloads enterprise engineering teams have been piloting for the past 18 months: automated code review, CI/CD pipeline agents, and customer-facing software that needs to take action rather than just generate text.

Where 3.5 Flash gives ground is on pure reasoning benchmarks. It scores 72.1% on ARC-AGI-2, which tests novel pattern recognition, versus 77.1% for Gemini 3.1 Pro. On Humanity’s Last Exam, it trails Pro 40.2% to 44.4%. Those gaps matter if your use case involves hard scientific reasoning or PhD-level knowledge synthesis. They matter less if your use case involves agents completing structured work across connected software systems — which describes the majority of enterprise agentic deployments currently generating revenue.

The SWE-Bench Pro result tells a more nuanced story. At 55.1%, 3.5 Flash narrowly beats Gemini 3.1 Pro’s 54.2%, but Claude Opus 4.7 leads that benchmark at 64.3% — a gap large enough that any team prioritizing production-quality software engineering at the repository level has a real reason to stay on Anthropic’s platform. The benchmark data does not produce a universal winner. It produces a routing map.

The Cost Argument Is the Real Story

Gemini 3.5 Flash is priced at $1.50 per million input tokens and $9.00 per million output tokens. Gemini 3.1 Pro runs $2.00 input and $12.00 output. GPT-5.5 is positioned considerably higher. Claude Opus 4.7 input costs roughly ten times what 3.5 Flash charges per token.

Google made the enterprise math explicit at I/O. The company’s stated claim: organizations processing around one trillion tokens per day could save over $1 billion annually by shifting 80% of workloads from other frontier models to 3.5 Flash. That figure is directionally plausible even if it requires assumptions about current provider mix and workload composition. For the chief technology officer reviewing an AI infrastructure budget line that has compounded sharply over the past two years, the sentence is hard to sit with unchallenged.

The switching cost argument has always been the counterweight to price competition in enterprise AI. Prompt engineering, fine-tuning, evaluation pipelines, and production integrations are genuinely expensive to migrate. But 3.5 Flash supports the same 1 million token context window and full multimodal inputs as its predecessors, and it runs on the same API surface developers are already using. For teams that have not built deep model-specific optimization, the migration calculus is lighter than it was a year ago.

Where the Independent Data Is Sparse

One important caveat: most of the headline benchmark numbers for Gemini 3.5 Flash come from Google’s own evaluation methodology. Independent validation from third-party organizations like Epoch AI or ARC Prize was not available at the time of this writing. The 78% SWE-Bench Verified figure — slightly different from the SWE-Bench Pro numbers — is the one score that reviewers have been able to corroborate against the benchmark maintainer’s public leaderboard. The agentic suite numbers are self-reported.

That is not unusual. Most model launches arrive with proprietary benchmark packages before the research community has produced independent evaluations. But it does mean enterprise procurement teams evaluating a migration should treat the MCP Atlas and Terminal-Bench numbers as directional claims, run their own internal evals on representative production workloads, and wait for Artificial Analysis or comparable independent benchmarkers to publish results before making irreversible platform decisions.

The Strategic Frame Is Bigger Than the Model

Gemini 3.5 Flash did not launch in isolation. At I/O 2026, Google simultaneously announced Gemini Spark — a 24/7 background agent that executes workflows across Gmail, Docs, Salesforce, and ServiceNow — and the Gemini Enterprise Agent Platform, a full-stack management layer for organizations running agent fleets at scale. The model launch and the enterprise platform launch are the same pitch: Google is not selling a better API. It is selling a vertically integrated AI operating layer, with Gemini 3.5 Flash as the inference engine running inside it.

That context matters for how to interpret the pricing. Cutting token costs 25% while outperforming a more expensive flagship on agentic benchmarks is a credible stand-alone value argument. But Google’s actual target is the enterprise software budget sitting behind the API bill — the Salesforce renewals, the ServiceNow contracts, the Microsoft 365 seat counts — all of which become easier to defend or harder to justify depending on whether the incumbent AI layer is outperforming Google’s stack on the tasks employees actually run. The enterprise software renewal cycle has become the most consequential battlefield in the AI platform war, and Gemini 3.5 Flash is Google’s most targeted weapon in that fight so far.

The model that earns the most revenue in enterprise AI is not the one that scores highest on every benchmark. It is the one that runs reliably enough, cheaply enough, and broadly enough to become the default infrastructure choice before competitors lock in the next contract cycle. Google’s 3.5 Flash is an explicit attempt to claim that position before OpenAI’s IPO — and the enterprise customer conversations it will generate — resets the room.

What to Watch Next

• Independent benchmark validation from Epoch AI, Artificial Analysis, or ARC Prize for Gemini 3.5 Flash’s MCP Atlas and Terminal-Bench 2.1 scores. Self-reported results are the current baseline; independent confirmation or revision will be the first real pricing signal for enterprise procurement teams evaluating a migration.
  
• Gemini 3.5 Pro release timeline. Google confirmed the Pro variant is in internal use and slated for the following month. Its benchmark profile — and whether it closes the gap with Claude Opus 4.7 on SWE-Bench Pro — will determine whether the 3.5 family is competitive at the top of the market, not just in the mid-tier routing decision.
  
• GitHub Copilot and Cursor adoption data for Gemini 3.5 Flash. Both platforms integrated the model at launch. Any developer survey or usage disclosure from either company through Q3 will be the first real-world signal on whether 3.5 Flash’s agentic benchmark advantage translates into production coding agent adoption at scale.
  
• OpenAI’s response. The company has not formally repositioned a product against 3.5 Flash’s pricing tier. Any announcement of a GPT-5.5-class model at a comparable or lower token cost would directly narrow the cost gap Google is using as its primary enterprise argument.
  
• Google Cloud Q2 2026 results. Alphabet reports in late July. Enterprise AI paid user growth and any Gemini Enterprise Agent Platform adoption commentary will show whether I/O 2026’s agentic positioning is converting to commercial contracts or remaining a developer-layer story with a longer enterprise sales cycle ahead of it.

On February 9, 2026, OpenAI activated paid advertising inside ChatGPT for free and Go-tier users in the United States. Six weeks later, the pilot had crossed $100 million in annualized revenue. By April, the company had opened a self-serve Ads Manager to any U.S. business. On June 6, Benji Shomair, OpenAI’s VP of Monetization, confirmed the pilot was live in the United Kingdom — the first market outside North America, Australia, and New Zealand — with Japan, South Korea, Brazil, and Mexico to follow. An advertising business that did not exist five months ago is now operating on four continents.

The pace is not an accident. OpenAI is heading toward a September IPO targeting a valuation above $1 trillion. The advertising business is part of the filing narrative whether it appears as a formal line item in the S-1 or not.

The Free Tier Problem That Ads Are Supposed to Solve

ChatGPT reached 900 million weekly active users in February 2026, roughly doubling its base from a year earlier. Of that audience, approximately 95 percent pay nothing. The paid subscriber base — Plus, Pro, Team, and Enterprise — accounts for around 50 million users. That ratio is the structural pressure the ad business is designed to address.

OpenAI is projecting $2.5 billion in ad revenue for 2026, rising to $11 billion in 2027, $25 billion in 2028, $53 billion in 2029, and $100 billion by 2030. Those projections, shared with investors and first reported by Axios in April, assume the platform reaches 2.75 billion weekly users by the end of the decade. The forecasts are not modest. Google’s total advertising revenue in 2025 was approximately $294 billion; OpenAI’s 2030 target represents about one-third of that, built from scratch, in four years.

Barclays published a separate projection in April that arrived at $102 billion by 2030 through a model built on user growth, query volume, and revenue per query. The methodology is different from OpenAI’s internal numbers; the destination is nearly identical. That alignment between an outside sell-side model and the company’s own investor presentations is what gives the projection real weight.

What “Intent Advertising” Is Actually Worth

The initial launch CPM of $60 — roughly three times what Meta charges on Facebook and Instagram — signaled how OpenAI was positioning the product from day one: not as a reach channel, but as a high-intent environment where users arrive with specific problems, not passive scroll habits. A person asking ChatGPT how to refinance a mortgage is more commercially actionable than a person encountering a mortgage ad in a news feed.

That framing is now being tested against a much wider advertiser base. The $200,000 minimum commitment that characterized the closed beta was eliminated in May when OpenAI opened its self-serve Ads Manager to any U.S. business, adding CPC bidding at a recommended starting rate of $3 to $5 per click. CPMs reportedly drifted down to $25 as the advertiser pool widened in April — a natural consequence of moving from curated brand partners to open access. The question now is whether the intent premium holds at scale or gradually compresses toward something closer to search-display blended rates.

“Creative variation has been a real key to success.”
— Benji Shomair, VP of Monetization, OpenAI, press roundtable, May 2026

The remark is revealing. In a high-intent environment, the context of a query shapes what an ad needs to do in ways that standard creative testing on Google or Meta does not surface. Someone researching running shoes mid-conversation requires a different creative execution than someone browsing a shopping feed. OpenAI is accumulating the behavioral data to act on that — but it is only four months into building that signal.

The Infrastructure Being Built in Real Time

Conversion tracking arrived in May. Cost-per-action bidding entered early access on June 5. Daily budget controls, DMA-level geo-targeting, custom audience targeting, and dynamic call-to-action units have all shipped since April. The build cadence is faster than anything Google or Meta attempted in their equivalent early periods — but both of those companies had years of advertising infrastructure elsewhere in their organizations before they launched the products that made them dominant.

OpenAI has hired quickly. David Dugan joined as Global Head of Ads in late March and confirmed Zalando as a UK launch partner. Technology integrations include Criteo, StackAdapt, Adobe, Kargo, and Pacvue. Shomair framed the access expansion explicitly around mission rather than margin: “Expanding access to advertising helps support our broader mission of making frontier AI more accessible to more people.” Whether investors read it as vision or as cover for a revenue necessity will depend on what the audited financials eventually show.

The measurement gap is the most material near-term risk. Early advertisers have reportedly complained about the absence of robust attribution tools, and the product feed ad format has moved slowly due to inconsistent merchant data quality. Those are solvable problems — Google’s first years were defined by similar limitations — but they constrain enterprise budget allocations while they persist. At $2.5 billion projected for 2026, any significant delay in measurement infrastructure arriving would leave the year-end number short.

What the Ad Business Changes About the IPO Math

OpenAI is projecting $14 billion in operating losses for 2026. Subscription revenue, enterprise licensing, and API fees are the current primary contributors. Advertising adds a structurally different revenue profile: high-margin, scaling with usage rather than new contract activity, and not dependent on compute provisioning in the way model inference is.

For the roadshow narrative, the advertising line matters more as a trajectory than as a current number. A platform with 900 million weekly users that generated $100 million annualized in six weeks, opened to self-serve in four months, and is now live in multiple international markets is a demonstrably functional ad business — not a pilot. How Goldman Sachs and Morgan Stanley frame that trajectory in the S-1 language will influence whether institutional investors treat the ad revenue projection as a credible second engine or as aspirational modeling. The Anthropic filing, which arrived first on June 1, will set the disclosure baseline that OpenAI’s prospectus has to meet or exceed on this exact question.

There is also a competitive dimension that the IPO framing will need to address. Google’s advertising revenue in 2025 reached approximately $294 billion. Meta earned close to $200 billion. Both companies have spent years building measurement infrastructure, advertiser relationships, and attribution pipelines that OpenAI is assembling in real time. The question is not whether ChatGPT can take share from those platforms — it is whether it can do so fast enough to matter to the revenue model before the company’s burn rate requires a resolution. The AI IPO sequencing between Anthropic and OpenAI means whoever discloses audited gross margins first sets the sector multiple. The advertising line, small as it is today, changes both the numerator and the story.

What to Watch Next

• The OpenAI public S-1 on SEC EDGAR. The first audited disclosure of advertising revenue as a formal line item — and whether the $2.5 billion 2026 projection is incorporated into the prospectus as formal guidance or presented as an internal estimate — will be the clearest signal of how central the ad business is to the IPO valuation case.
  
• CPM stability through Q3. The drop from $60 to $25 as the advertiser pool widened is worth tracking. Whether pricing stabilizes at self-serve levels or continues to compress will determine the actual revenue contribution and whether the Barclays $102 billion model holds its assumptions.
  
• Japan, South Korea, Brazil, and Mexico activation dates. Each market adds to the addressable user base for advertisers. The pace of international rollout through Q3 will be the most visible indicator of whether the self-serve infrastructure can support simultaneous multi-market builds.
  
• Measurement and attribution tool launches. Cost-per-action bidding entered early access on June 5. Full rollout and any disclosure of conversion rates against Google and Meta benchmarks would give enterprise advertisers the comparison data they need to commit significant budget — and give analysts a basis for adjusting the 2027 projection.
  
• Google and Meta response. Neither company has formally repositioned a product against ChatGPT ads. Any announcement of a conversational ad format from Google or an AI-native intent product from Meta would directly challenge OpenAI’s pricing premium and force a reassessment of the long-term market share assumptions underlying the 2030 projections.

It is now over. On June 3, Reuters reported that DeepSeek is preparing to raise approximately 50 billion yuan, or $7.4 billion, in its first external funding round. Tencent and CATL are expected to be the largest outside investors, committing around 10 billion yuan and 5 billion yuan respectively. Liang himself is reportedly injecting 20 billion yuan of personal capital. IDG Capital and Monolith Capital are also among the prospective investors. The round could value DeepSeek between $52 billion and $59 billion post-money — a figure that, as recently as April, analysts were placing closer to $10 billion.

Why a Hedge Fund Can No Longer Do This Alone

The High-Flyer model was genuinely elegant. Liang built algorithmic trading infrastructure in the mid-2010s that proved directly transferable to training large language models. The compute clusters High-Flyer built for quantitative finance became DeepSeek’s early GPU backbone. When DeepSeek formally incorporated in July 2023, the capital was already in-house, the hardware was available, and the team had no obligation to any LP or VC timeline. DeepSeek trained its V3 model for a reported cost of around $5.6 million — roughly five percent of what GPT-4 cost to train — and released the weights open-source.

That efficiency story was always partly a function of constraints. DeepSeek has been blocked from accessing Nvidia’s highest-end accelerators since the U.S. expanded export controls. Its V3 model was trained on H800 chips, the first Nvidia product specifically designed to fall outside U.S. export restrictions, before those too were restricted. The V4 family, launched April 24, 2026 with 1.6 trillion total parameters across V4-Pro and V4-Flash, was optimized to run on Huawei Ascend silicon — a deliberate signal that DeepSeek is engineering around the chip supply wall rather than waiting for it to come down. NIST’s Center for AI Standards and Innovation evaluated V4’s capabilities against frontier models and found it lags the best Western closed models by approximately eight months, a narrower gap than most industry observers expected given the hardware constraints.

The problem is that closing that gap further requires compute at a scale that no hedge fund’s annual returns can sustain. Training runs at the frontier now consume infrastructure investments that make DeepSeek’s previous $5.6 million efficiency story structurally impossible to repeat. The funding round is a recognition that the physics of scale have caught up with even the most capital-efficient lab in the industry.

The $59 Billion Valuation Is a Strategic Statement, Not Just a Price

The valuation trajectory is itself the news. As recently as May, reporting placed DeepSeek’s implied valuation in the $10 to $20 billion range. The $52–59 billion post-money figure that emerged in Reuters’ June 3 reporting represents a tripling of expectations within weeks. That is not the result of a business model disclosure or revenue revelation — DeepSeek has made no public communication about its revenue, and Liang has explicitly told prospective investors the company will prioritize fundamental research over near-term commercialization. The valuation is being set entirely by strategic scarcity: there is one credible open-source AI lab operating at frontier scale outside the U.S., and investors are pricing access to it before the window closes.

The investor mix reinforces that read. Tencent and CATL are not venture capital funds making financial return calculations. Tencent needs DeepSeek proximity to stay competitive with Alibaba’s Qwen platform. CATL, primarily known as the dominant EV battery supplier, has been expanding into AI data center infrastructure, and a DeepSeek relationship gives it a demand anchor for that push. The government-backed National Artificial Intelligence Industry Investment Fund and IDG Capital are institutional signals that Beijing views this round as a strategic asset consolidation, not simply a private market transaction. The Western investor category is notably absent from all current reporting on the round.

What the No-Revenue Pitch Actually Means for Investors

Liang’s messaging to investors is unusual at a moment when every other frontier AI lab is under pressure to demonstrate a path to profit. OpenAI restructured its corporate governance and filed for a $1 trillion IPO in part to prove it can convert its 600 million user base into sustainable revenue. Anthropic raised $65 billion in May on a reported gross margin improvement to above 70%. The capital cycle that has driven AI valuations to generational extremes has been sustained, in part, by revenue and margin disclosure — something DeepSeek has systematically withheld.

Liang’s pitch is structurally different: fund us to pursue AGI and keep open-source models flowing, and the commercial return comes from being indispensable to Chinese AI infrastructure rather than from subscription revenue. That argument worked internally while High-Flyer’s returns covered the bills. It is now being tested against a roster of strategic investors who have their own board obligations, earnings calls, and competitive pressures. The question is not whether Liang means it. It is whether Tencent’s and CATL’s governance structures can hold that commitment across multiple quarters once compute costs and competitive intensity rise further.

The Hardware Wall Is the Thesis, Not the Footnote

The most important underreported dimension of this round is what DeepSeek cannot spend the capital on. Nvidia’s H100 and Blackwell architectures are both barred from export to China. The H20, a downgraded chip Nvidia designed specifically for the Chinese market, was added to the restricted list in 2025. DeepSeek’s V4 family was trained on Huawei Ascend 950PR chips, a decision that required months of software migration from Nvidia’s CUDA framework to Huawei’s CANN stack. That migration cost was, in effect, the price of hardware independence.

A $7.4 billion round going into a lab that cannot access the leading GPU architecture creates a distinct investment thesis from anything in the Western AI stack. The capital will fund Huawei Ascend infrastructure, Cambricon silicon deployment, domestic data center expansion, and the engineering talent required to optimize models for non-Nvidia hardware at scale. If that work succeeds, it validates the export control framework’s core weakness: the assumption that hardware constraints are permanent rather than workable. If it fails to close the eight-month capability gap that NIST measured, the round becomes a strategic investment in a geopolitical asset rather than a commercial technology bet — and that distinction matters to how investors should model their exit.

The round is expected to close within weeks, per Reuters. Discussions remain in flux and terms could still change. But the fact that Liang is raising at all — after three years of deliberate independence — is the answer to the question that matters most. Even the most capital-efficient lab in the world has hit the scaling wall.

What to Watch Next

• Round close confirmation and final investor list. The sub-10-investor structure means any addition of a state fund or strategic outside the current roster would shift the geopolitical read on the deal significantly. Watch for Alibaba, which has been reported as an interested party, and whether it participates despite competing with DeepSeek through its own Qwen platform.
  
• Huawei Ascend 950PR production capacity. DeepSeek’s ability to deploy the capital depends entirely on Huawei’s ability to manufacture chips at scale without Nvidia TSMC wafer access. Any supply constraint announcement from Huawei will directly limit DeepSeek’s compute ramp timeline.
  
• DeepSeek V4 stable release date. The April 24 launch carried a “preview” designation. A stable release with confirmed production readiness would be the first signal that the V4 family is generating the usage and API revenue that would eventually require commercialization disclosure to investors.
  
• Any SEC or CFIUS review of Western investors participating in the round. No Western institution is currently reported to be involved. If one surfaces, U.S. regulatory scrutiny of the transaction becomes a material variable for both the investor and the round timeline.
  
• Tencent’s next earnings commentary on AI infrastructure spend. Tencent’s AI investment posture — whether it positions the DeepSeek stake as a strategic partnership or a financial investment — will be the first public disclosure of the relationship’s commercial terms.

The enterprise hardware tier — the servers, networking gear, and storage systems that sit beneath the GPU stack and handle the actual deployment of AI into corporate workflows — just posted its clearest earnings cycle yet. Dell, HPE, and NetApp all reported within 72 hours of each other. The numbers are not incremental. They are structural.

Dell Set the Tone First

Dell Technologies reported fiscal first-quarter 2027 results on May 28, 2026, and the figures were difficult to contextualize at normal scale. Total revenue reached $43.8 billion, up 88% year over year — the fastest growth rate in the company’s history as a public company. AI-optimized server revenue hit $16.1 billion in a single quarter, a 757% year-over-year increase. The company booked $24.4 billion in AI orders and exited the quarter with a record $51.3 billion in AI backlog.

Vice Chairman and COO Jeff Clarke raised the company’s full-year AI server revenue forecast to $60 billion. To put that in context: Dell’s entire revenue for fiscal year 2024 was $88 billion. It is now guiding for more than two-thirds of that figure from a single product category in a single year.

“Demand was even stronger than revenue growth. Orders more than doubled, significantly outpacing revenue, resulting in a record company backlog. Customer investments in agentic AI and AI inferencing accelerated.”
— Antonio Neri, President & CEO, Hewlett Packard Enterprise, Q2 FY26 Earnings Call, June 1, 2026

HPE: Where Demand Runs Ahead of Revenue

HPE’s fiscal second-quarter 2026 results, reported June 1, reinforced the Dell read from a different angle. Revenue of $10.7 billion was up 40% year over year, and non-GAAP EPS of $0.79 increased 108% — but the more important number was in the backlog. Orders more than doubled and significantly outpaced revenue growth, leaving HPE with a record AI systems backlog of $5.9 billion as it entered the third quarter. The company booked $1.8 billion in new AI systems orders during the quarter alone, bringing cumulative AI systems bookings to $16.4 billion.

The supply constraint that is limiting HPE’s revenue conversion — memory and processor availability — is the same one Dell flagged. In both cases, the message is the same: the ceiling on near-term revenue is supply, not demand. That distinction matters. A demand cliff can emerge without warning. A supply constraint resolves as procurement agreements are locked in and capacity comes online, and HPE noted it already has supply locked for known demand scenarios.

HPE raised its full-year fiscal 2026 EPS guidance by more than 40%, to $3.35–$3.45, and its free cash flow outlook to at least $3.5 billion. It also issued an early fiscal 2027 framework — a step the company rarely takes — citing durable demand visibility as the reason. CEO Antonio Neri was specific about what is driving the acceleration: agentic AI workloads. The shift from training to inference and, increasingly, to agentic deployment is expanding the addressable hardware market beyond pure GPU infrastructure into networking, compute, storage, and memory simultaneously.

NetApp Is the Storage Signal Everyone Is Underreading

NetApp reported fiscal fourth-quarter 2026 results on May 28. Total revenue of $1.95 billion was up 12% year over year — modest by Dell and HPE standards, but the composition matters more than the headline. All-flash array revenue for the quarter hit $1.2 billion, up 18% year over year, and the full-year all-flash figure reached $4.2 billion. CEO George Kurian disclosed more than 1,100 AI and data preparation wins for the fiscal year, compared with roughly 400 the prior year.

Kurian’s explanation of the storage demand dynamic is worth tracking closely. Roughly half of NetApp’s AI use cases were tied to data preparation and large-scale analytics — not model training or inference directly, but the underlying data work that precedes both. The other half was split between training and inferencing. That breakdown reveals something about where enterprise AI deployment actually lives: in messy, unstructured data pipelines that require high-performance, scalable storage infrastructure long before a GPU ever processes a query.

NetApp guided fiscal 2027 revenue growth at roughly 8% at the midpoint, raised its buyback authorization by $1 billion, and stated its intent to return up to 100% of free cash flow to shareholders. For a company trading at more restrained multiples than its server-focused peers, the shareholder return posture signals management confidence in the durability of the demand environment.

Why This Cluster of Results Changes the Portfolio Conversation

The hyperscalers — Microsoft, Google, Amazon, Meta — have been the visible faces of the AI infrastructure buildout. Their capital expenditure commitments have been covered in detail here and elsewhere. What this week’s earnings cycle revealed is that the $690 billion in hyperscaler capex is not an abstraction. It is flowing through the hardware supply chain as real purchase orders, real backlog, and real revenue at companies that trade at fractions of the hyperscaler multiple.

Dell’s Infrastructure Solutions Group now generates more revenue in a single quarter than most large-cap companies generate in a year. HPE’s backlog-to-revenue gap is a leading indicator of near-term earnings revisions, not a risk factor. NetApp’s AI win count tripled year over year. Taken together, these three reports are the earnings-cycle confirmation of a thesis that has been priced speculatively in the hyperscalers for two years and is only now showing up in the enterprise hardware tier at scale.

The market moved accordingly. Dell surged more than 30% after hours on May 28. HPE jumped roughly 30% in after-hours trading on June 1. NetApp moved in the same session. These are not typical earnings reactions. They reflect a rerating event — investors updating their models on what AI infrastructure actually looks like when it converts from capital commitment to shipped revenue.

What the Margin Picture Is Telling You

Not every signal here is straightforwardly positive. Dell’s AI server gross margins are compressing. The Infrastructure Solutions Group margin came in at 10.5% in the most recent quarter, below historical levels, because AI-optimized servers carry lower gross margins than traditional enterprise hardware. The volume is extraordinary; the per-unit profitability is thinner. HPE has flagged that it expects continued elevated costs into fiscal 2027 due to supply chain pressures, particularly in memory. NetApp is managing component cost increases affecting its all-flash line.

The margin question is the right analytical tension to hold. Revenue growth at 40% to 88% is extraordinary. The durability of that growth at current margins — or the pace at which margins normalize as supply constraints ease and mix shifts toward software and services — is the variable that separates this from a one-cycle event. HPE’s explicit commentary on its path to margin recovery, tied to inferencing demand growth through the end of the decade, suggests management sees the margin compression as transitional rather than structural. The next two quarters will either confirm or challenge that read.

What to Watch Next

• HPE Q3 FY26 results, expected late August — the company guided revenue of $11.5 billion to $12.1 billion. Whether backlog conversion accelerates as supply constraints ease is the central test. Any commentary from CFO Marie Myers on whether the $5.9 billion AI systems backlog is converting at pace or continuing to build will reset near-term estimates.
  
• Dell Q2 FY27 results, also expected late August — the $60 billion full-year AI server revenue target implies roughly $15 to $16 billion per quarter for the remaining three periods. Whether Dell hits the midpoint is the single clearest test of whether the AI server supercycle is sustaining or beginning to exhaust addressable near-term demand.
  
• Memory pricing and supply availability. Both Dell and HPE explicitly cited memory as the primary near-term supply constraint. Any announcements from Micron, Samsung, or SK Hynix on HBM and NAND allocation timelines will directly affect how quickly either company can convert backlog into revenue — and whether Q3 guidance can be exceeded or will face continued supply-side friction.
  
• AI inferencing margin data as it surfaces. HPE’s Neri said he expects inferencing to represent the majority of long-term AI demand. The margin profile of inferencing hardware — which requires less custom GPU and more networking, memory, and storage — is where HPE and NetApp are best positioned relative to Dell. Watch for any product-level margin disclosures that begin to separate inferencing economics from training economics in public filings.
  
• Enterprise AI adoption data from software companies. HPE and NetApp’s backlog growth is a forward indicator, but the sustainability of that demand ultimately depends on whether enterprises are deploying AI at scale rather than piloting it. Salesforce, ServiceNow, and SAP earnings will provide the best read on whether enterprise AI spend is compounding or flattening at the application layer.

The sequencing is not symbolic. Under SEC rules, a public prospectus must be filed at least 15 days before the roadshow begins. The company that files its public S-1 first sets the financial disclosure terms that every analyst covering the second company will use as their baseline. If Anthropic’s gross margins and cost structure reach the market before OpenAI’s numbers are public, Anthropic controls the comparable. If OpenAI’s audited financials land first at a September roadshow, OpenAI sets the multiple.

“This gives us the option to go public after the SEC completes its review.”
— Anthropic, official statement, June 1, 2026

The Revenue Crossover That Rewrote the Narrative

Anthropic’s Series H announcement at the end of May disclosed a revenue run rate of $47 billion, up from $30 billion in April and $9 billion at the end of 2025. That growth curve, which Inc. and CNBC both confirmed from the company’s fundraising disclosures, represents a pace with few comparables in enterprise software history. OpenAI’s run rate, by contrast, is estimated by analysts at roughly $36 billion as of May, based on the company’s reported $2 billion monthly revenue trajectory. Anthropic has publicly surpassed its primary rival on this metric for the first time.

The reversal matters for the IPO race because run-rate revenue, however imperfect as a proxy for real earnings, is how institutional allocators are currently sizing these deals. A company that enters the roadshow with a higher disclosed run rate commands the anchor comparison, and Anthropic has now built a lead it can sustain through whatever window the SEC review opens. According to CNBC, the revenue growth has been driven primarily by enterprise adoption of Claude for coding and agentic workflows, with Claude Code emerging as the product most directly responsible for the acceleration.

What the Confidential Process Conceals — and When It Stops

Both companies are currently inside the SEC’s confidential review process, which keeps audited financials, cost structures, customer concentration, and risk factors sealed from public view. For Anthropic, the most consequential number still unverified is its gross margin. The company has disclosed a run-rate improvement from roughly 38% to above 70%, a figure that has circulated in investor materials and secondary reporting. But that claim has not been tested against an audited income statement. The S-1 is where the math becomes checkable.

For OpenAI, the sealed document contains something different: loss disclosures. Multiple outlets, including Fortune and a detailed analysis citing the company’s own internal figures, have reported that OpenAI loses approximately $1.22 for every dollar of revenue generated. At a $36 billion run rate, that implies annual losses in the range of $14 billion. How OpenAI’s bankers — Goldman Sachs, Morgan Stanley, and JPMorgan Chase — frame that burn rate in the prospectus will define the risk section that every institutional allocator reads first. Sam Altman acknowledged internally, per reporting by The Information, that filing for an IPO is different from being ready to go public.

The Timeline Math Is Tighter Than It Looks

SEC review of confidential filings typically runs 60 to 90 days, with multiple rounds of comment letters. OpenAI filed May 22. That puts its earliest realistic public S-1 in late August, with a September roadshow and pricing still achievable but requiring a clean first-round review with no material accounting questions. Anthropic filed June 1. Its public S-1 could arrive as early as late August as well, if the SEC’s workload and comment calendar allow both reviews to run in parallel. The 10-day difference in filing dates may compress to near-zero by the time both prospectuses are public.

That compression sets up a scenario the market has not seen before: two companies with near-$1 trillion implied valuations releasing detailed, audited financial disclosures within weeks of each other, after a decade of operating entirely outside public reporting requirements. Axios has reported that both companies are targeting a listing window between Labor Day and Thanksgiving, and that they are working with many of the same banks. The roadshow period, when it arrives for each company, will be the first moment retail and institutional investors can access numbers that have previously been available only to late-stage private market participants.

SpaceX’s June 12 Debut Sets the Tone for Everything Else

Before either Anthropic or OpenAI reaches a roadshow, SpaceX prices on June 12. The deal is targeting up to $75 billion at a valuation of $1.75 to $1.8 trillion, which would make it the largest IPO in history. Its first-day trading performance and the quality of the institutional order book will function as a real-time stress test of whether public markets in 2026 will absorb technology companies at these multiples. A strong open confirms that the rate environment and investor appetite can support what comes next. A soft open compresses the September window for OpenAI and could push Anthropic into Q4 or beyond, regardless of how clean its SEC review process turns out to be.

The 10-year Treasury yield, which touched 4.6% during May’s bond-market volatility, is the rate variable the underwriters are watching most closely. The June 16-17 FOMC meeting under Chair Kevin Warsh will be the first policy signal in the post-SpaceX pricing environment. Any language suggesting rate hikes are back under consideration would immediately reprice the revenue-multiple math that all three valuations depend on. At 95 times revenue and a net loss, these are not interest-rate-neutral deals.

The Comparable Problem Neither Company Wants

Once both prospectuses are public, analysts will have, for the first time, two directly comparable AI frontier-lab financials to work from. That comparison will not favor whichever company has the weaker gross margin, higher customer concentration, or less defensible cost structure. Both companies have operated for years with the benefit of no public comparable. The SpaceX prospectus has already disrupted that dynamic at the infrastructure layer. Anthropic’s and OpenAI’s S-1s will do it at the model layer, and neither company controls how that comparison lands once the numbers are visible side by side.

The sequencing race is really a narrative race. Filing first, disclosing stronger unit economics, and pricing into a stable rate environment are the three variables that determine which company walks away from the fall IPO window having set the terms. Anthropic moved on June 1 to make sure it has a shot at all three.

What to Watch Next

• SpaceX first-day trading, June 12. Institutional appetite at the open and the closing price relative to the $1.75–$1.8 trillion anchor will be the most direct signal of whether the rate environment can support the two AI IPOs that follow.
  
• The FOMC meeting, June 16-17. Chair Warsh’s first statement as Fed chair. Any language reintroducing rate-hike risk would immediately reprice the revenue multiples underpinning both AI listings and could force timeline slippage into Q1 2027.
  
• Anthropic’s public S-1 on SEC EDGAR. The audited gross margin figure is the single most consequential number in the filing. Whether it confirms or revises the reported 70%-plus improvement from 38% will reset private-market valuations across the AI sector the day it publishes.
  
• OpenAI’s public S-1 and loss disclosure framing. How Goldman Sachs, Morgan Stanley, and JPMorgan structure the burn-rate narrative — particularly whether they position current losses as a short-term infrastructure investment or a structural characteristic of the business — will determine the multiple institutional investors will accept.
  
• The order of roadshows. If Anthropic begins its roadshow before OpenAI’s S-1 is public, Anthropic controls the sector valuation conversation. If OpenAI prices first, it sets the comparable Anthropic’s bankers must defend. Watch for any underwriter communication that signals which company intends to move faster through the SEC comment process.

The headline numbers from Q1 were solid but unspectacular: revenue of $177.8 billion, up 7.3%, and adjusted operating income growth of roughly 5% in constant currency. What was harder to see in the top-line figures was a structural shift in where Walmart’s profits are actually coming from — and what that shift implies about the return on the company’s AI and automation build-out.

The High-Margin Layer Is Growing Faster Than the Core

Advertising revenue at Walmart U.S. grew 36% in Q1, with the Walmart Connect platform up 44% excluding the VIZIO integration. Membership fee revenue grew 17.4% globally, with U.S. Walmart+ fee revenue hitting a record for a fiscal first quarter. Marketplace net sales in the U.S. grew nearly 50%. Taken together, advertising and membership fees now account for roughly one-third of consolidated operating income — a figure that would have been difficult to model five years ago for a company whose identity was built on low-margin grocery volume.

The connection to AI and automation is not incidental. Walmart’s advertising business runs on data — specifically, its ability to close the loop between a digital ad impression and a physical or online purchase at a scale that Google and Meta cannot replicate with their own retail infrastructure. The more automation improves inventory visibility and order accuracy, the more valuable the first-party data becomes, and the more advertisers will pay for placement inside the Walmart ecosystem.

Automation Is Hitting the Numbers Rainey Promised

CFO John David Rainey said on the Q4 FY2026 earnings call that technology-enabled productivity benefits are critical to the company’s ability to grow its core business at lower marginal cost. The Q1 data shows that thesis is tracking, even against a $175 million fuel cost headwind that knocked roughly 250 basis points off operating income growth.

“Our teams are adopting innovative technologies, driving productivity through automation and growing higher-margin commerce solutions.”
— John Furner, President and CEO, Walmart Inc., Q1 FY2027 Earnings Release, May 21, 2026

About 60% of Walmart U.S. stores are now receiving freight from automated distribution centers, and roughly half of the company’s e-commerce fulfillment center volume is automated. Twenty-three of the company’s 42 regional distribution centers in the U.S. are being retrofitted. Walmart has described this two-year window as the peak of its annual capital spending on supply chain automation and store remodels. The expected return on that peak is operating income growing faster than sales — which Q1 delivered, even masked by fuel costs, and which Q2 guidance implies more explicitly: operating income growth of 7% to 10% against net sales growth of 4% to 5%.

The Metric Investors Should Be Watching in August

The Q2 guidance range — operating income growth of 7% to 10% against net sales growth of 4% to 5% in constant currency — is the clearest single number Walmart has given for whether the AI-driven model is generating return at scale. Adjusted EPS guidance of $0.72 to $0.74 implies the company expects to deliver operating leverage even as fuel costs are expected to remain elevated. Rainey said on the Q1 call that the Q2 operating income guidance is the best the company has given in a decade and a half.

The cleanest test will come from two specific lines in the August 20 print: gross profit rate direction and the advertising revenue growth number. Gross profit rate improved 6 basis points in Q1, aided by merchandise mix and the continued shift toward higher-margin categories and business lines. If that improvement holds or accelerates into Q2, it confirms that the mix shift is durable, not a one-quarter feature of favorable category trends. If advertising growth holds near 30% or above for the third straight quarter, it validates that the data infrastructure underlying the ad business is compounding, not plateauing.

What AI Investment Looks Like in Practice at This Scale

Walmart’s AI deployment is not a single product or a pilot program — it is woven into the operating model across three distinct functions. The first is supply chain: automated distribution centers reduce labor intensity in freight processing and improve inventory accuracy, which Furner noted has measurably reduced markdowns. The second is merchandising: the Wally AI agent assists buyers with out-of-stock and overstock root-cause analysis, reducing reactive ordering decisions that erode margin. The third is customer-facing revenue: Sparky, the AI assistant in the Walmart app, drives engagement that flows through to advertising and Walmart+ membership conversion.

This is distinct from the enterprise software AI story — the one playing out at Salesforce and ServiceNow, where companies are selling AI access on top of existing SaaS contracts. Walmart’s AI investment is embedded in the cost and revenue structure of the world’s largest retailer, and its returns are measured in basis points on a $177 billion quarterly revenue base, not in software ARR. That makes the numbers harder to isolate and easier for the market to miss — which is precisely why the August print deserves closer attention than it usually gets from investors whose Walmart position is a defensive allocation.

The question the Q2 print will answer is whether the operating leverage story Walmart has been building toward is arriving on the timeline its capex implied — or whether fuel, labor cost inflation, and tariff-related uncertainty are compressing the window. Based on the Q1 read-through and the guidance Rainey put on record, Walmart’s management believes the window is open. The tariff environment adds an additional test for any retailer with a global supply chain, but Walmart’s pricing power and supplier relationships give it more room to absorb or pass through cost changes than most competitors at its scale.

What to Watch Next

• Walmart Q2 FY2027 results, August 20 — the operating income growth number against the 7%–10% guided range is the primary test. Any delivery at the top of that range would be the strongest evidence yet that the automation-and-mix model is compounding. A miss below 7% would force a re-examination of whether fuel costs and tariff pressure are structural headwinds, not temporary ones.
  
• Advertising revenue growth rate in Q2. At 37% globally in Q1, the business is the fastest-growing major segment Walmart operates. Watch for whether growth continues above 30% or begins to moderate as the base of comparison expands. Deceleration below 25% would signal the easy gains from audience scaling are behind them and the harder work of monetization efficiency is ahead.
  
• Gross profit rate in Q2. A second consecutive quarter of year-over-year improvement in gross profit rate, driven by merchandise mix, would be one of the more durable signals that the technology-enabled shift toward higher-margin categories is taking hold at the unit level, not just in the consolidated income statement.
  
• Walmart+ net add disclosure. Membership fee revenue grew double digits in Q1, setting a record for the quarter. Whether the absolute subscriber count is growing or whether revenue growth is being driven primarily by price and mix will clarify how much runway remains before the membership business requires a structural change.
  
• Capital expenditure guidance revision. Walmart has guided capex at approximately 3.5% of net sales for FY2027, describing this as the peak spending window. Any update to that guidance — upward or downward — would reset the model for when operating cash flow begins to recover materially from the current investment cycle’s peak.