Two years ago, Waymo’s commercial footprint was essentially one city — Phoenix, Arizona — where the flat, sun-drenched streets and well-mapped suburban grid made for a forgiving operational environment. Today the company operates driverless fleets in San Francisco, Los Angeles, Phoenix, Austin, Atlanta, and Washington D.C., with Miami on track for commercial launch by Q3 2026.
The fleet has grown to approximately 2,400 active vehicles, predominantly fifth-generation Jaguar I-Pace SUVs retrofitted with Waymo’s proprietary sensor stack, with the company’s purpose-built Zeekr-platform vehicles beginning to enter rotation in San Francisco this quarter. Fleet expansion is now pacing at roughly 200 new vehicles per month, up from 50–70 per month in mid-2024.
Average weekly trips per active vehicle have climbed to 85, compared to the roughly 60 industry analysts considered the break-even threshold for unit economics. Waymo has not disclosed per-vehicle revenue figures, but Alphabet CFO Anat Ashkenazi confirmed in Monday’s earnings call that Waymo’s revenue run-rate is “tracking toward $1 billion annually” — the first time Alphabet has offered a concrete revenue figure for the unit.
Perhaps more significant than the ride count is the accumulating safety record. Waymo published updated autonomous driving safety data last week showing its vehicles are involved in injury-causing collisions at a rate approximately 6.7 times lower than the average human driver in comparable urban environments, based on 50 million autonomous miles of operation.
That data — independently audited by the Swiss Testing Institute — is increasingly cited by municipal regulators weighing whether to permit commercial robotaxi operations. San Jose and Denver have both opened licensing processes in the past 60 days, citing Waymo’s safety disclosures as the triggering evidence.
The safety moat is hard to replicate quickly. Tesla, which has promised a paid robotaxi launch for years, has not yet received commercial operating authority in any US jurisdiction for fully driverless (no safety driver) operations. Zoox, Amazon’s robotaxi unit, is operating a closed campus shuttle service in Foster City but has not announced a public commercial timeline.
For years, critics of the autonomous vehicle industry argued that the economics would never work — that insurance liability, sensor costs, and remote monitoring overhead would make robotaxis permanently more expensive than human drivers. That argument is losing ground.
Waymo’s sensor hardware cost per vehicle has dropped by an estimated 70% since the launch of its fourth-generation system in 2021, according to supply chain analysis from Canaccord Genuity. Remote monitoring, which once required near-one-to-one human oversight ratios, now runs at roughly one operator per 20 vehicles during standard conditions, and one per 50 during high-confidence autonomous segments.
The company has been tight-lipped on when it expects to reach profitability at the unit or segment level. But with Alphabet committing to continued Waymo investment — the parent company has injected over $11 billion into the unit since 2009 — and Waymo’s own external funding round of $5.6 billion closed in late 2024, the runway is not the constraint it once was.
The 10 million ride milestone is, at its core, a signal to the rest of the industry: the question for autonomous vehicles has shifted from “will this ever work?” to “who gets to scale it?”
Sources: Alphabet Q1 2026 Earnings; Waymo Safety Report April 2026; Canaccord Genuity AV Hardware Cost Analysis 2026
]]>AMD’s MI350X, which began shipping to hyperscaler customers in volume during Q4 2025, delivers an estimated 35% improvement in large language model inference throughput compared to its predecessor. More critically, it has closed the software ecosystem gap that long hampered AMD’s pitch to enterprise buyers. ROCm 7.0, AMD’s GPU computing platform, now supports the full PyTorch and JAX model training stack with near-parity performance to NVIDIA’s CUDA in several production benchmarks.
Meta, Microsoft Azure, and Oracle Cloud have all confirmed MI350X deployments for inference workloads in 2026 — a roster that would have drawn skepticism as recently as late 2024. AMD CEO Lisa Su said on Monday’s earnings call that the company has “line of sight to over $20 billion in AI accelerator revenue” for full-year 2026, up from a prior target of $15 billion set in January.
AMD’s AI chip market share has grown from roughly 9% in early 2025 to an estimated 14–16% in Q1 2026, according to industry analysts at Mercury Research and Omdia. NVIDIA still controls approximately 75–78% of the AI accelerator market, with the H200 and upcoming Blackwell Ultra maintaining dominant positions in training workloads where CUDA’s software depth remains decisive.
The gap remains wide, but the direction of travel matters. Enterprise procurement teams, under pressure to reduce single-vendor dependency after NVIDIA’s persistent allocation constraints through 2024 and early 2025, have been actively qualifying AMD hardware as a second-source option. Several Fortune 500 companies confirmed to TechPulse that MI350X has moved from “evaluation” to “production” status in their infrastructure plans.
AMD’s gross margin expanded to 54.2% in Q1, up from 51.8% a year ago, reflecting the higher ASPs commanded by AI accelerators relative to the broader processor portfolio. The company posted $1.6 billion in net income on $8.3 billion in total revenue.
The strategic story for AMD is less about hardware specifications — where the gap with NVIDIA is now measured in percentages rather than multiples — and more about ecosystem maturity. For the past three years, NVIDIA’s CUDA platform has been the primary reason enterprises have accepted long wait times and premium pricing for H100 and H200 allocations.
AMD has invested aggressively in ROCm and has partnered with cloud providers to pre-configure MI350X instances with optimized inference stacks for common workloads including LLM serving, image generation, and multimodal model deployment. The company also acquired Mipsology, a model optimization startup, in late 2025 to accelerate quantization and sparsity support on its hardware.
Investors responded positively, with AMD shares rising approximately 4% in after-hours trading. The company guided Q2 2026 revenue to $8.7–9.1 billion, ahead of consensus estimates of $8.4 billion.
NVIDIA reports its own Q1 2026 results next month. With AMD narrowing the gap and custom silicon from Google, Amazon, and Microsoft eating into the hyperscaler segment, the AI chip market is entering a period of genuine competition — one that enterprise buyers, after years of constrained supply, are actively welcoming.
Sources: AMD Q1 2026 Earnings Release; Mercury Research AI Accelerator Market Share Q1 2026; Omdia Semiconductor Intelligence
]]>The shift from hype to hard data is driving a new cycle of institutional commitment. Eli Lilly, AstraZeneca, and Novartis have each signed multi-year AI research partnerships worth between $200 million and $700 million with platform companies in the last 18 months. The industry is no longer asking whether AI can accelerate early-stage discovery — it is now pricing in the assumption that it will.
Isomorphic Labs, the Google DeepMind spinout that emerged from the AlphaFold breakthroughs, has more than a dozen drug candidates in active development across oncology and rare disease indications. Recursion Pharmaceuticals, publicly listed and operating a platform that screens tens of millions of drug-disease combinations weekly, now has multiple programs in Phase II trials. Insilico Medicine, which drew attention in 2023 by becoming the first company to advance an entirely AI-designed molecule into human trials, has expanded its pipeline to over 30 programs.
The common thread is speed. Where traditional hit identification might take 18 to 36 months, AI platforms are compressing that window to three to six months in documented cases. For large pharma, where R&D productivity has been declining for years, that is not an incremental improvement — it is a structural shift in how early development can be financed and de-risked.
Analysts at Evaluate Pharma now estimate the AI drug discovery market at approximately $4 billion in contracted revenue for 2026, up from under $1 billion in 2022. Venture funding into the sector has remained elevated even as broader biotech financing tightened, with specialized AI pharma companies raising over $3.5 billion collectively in 2025.
The productivity argument is also gaining regulatory backing. The FDA’s Center for Drug Evaluation and Research issued updated guidance in early 2026 acknowledging AI-generated preclinical evidence in Investigational New Drug applications, a signal that the agency is building the framework to evaluate these submissions on their merits rather than their origin. The European Medicines Agency is developing parallel guidance expected later this year.
Recursion’s CEO has publicly cited a 40% reduction in average time-to-candidate across the company’s recent programs, a figure that, if reproducible at scale, would represent one of the most significant efficiency gains in pharmaceutical history.
The optimism is not without caveats. Phase II success rates in oncology remain stubbornly low regardless of how candidates are identified — AI does not yet change the biology of what happens in humans. Several early AI-discovered molecules have failed in mid-stage trials, and critics argue that compressed discovery timelines simply move the failure point earlier and cheaper, rather than fundamentally improving it.
There is also a data access problem. The best AI drug discovery platforms require proprietary biological datasets that only the largest pharma companies and specialized research hospitals possess. Smaller biotechs without those assets are building on public databases that may not reflect the patient populations they are targeting.
Still, the industry is past the stage where these arguments halt investment. With multiple programs now in Phase II and one approaching Phase III readiness, 2026 is shaping up as the year AI drug discovery stops being a future story and starts generating present-tense outcomes.
]]>According to figures compiled by Gartner in its Q1 2026 security spending report, enterprise allocation to AI-native security platforms grew approximately 43% year-over-year, outpacing total security spending growth of 14%. The divergence reflects a straightforward calculation: traditional signature-based and rules-based systems cannot respond at the speed or volume of attacks now being launched using large language models and automated exploitation frameworks.
Security researchers at several major firms documented a sharp increase in 2025 in what they categorize as AI-augmented phishing campaigns — messages that are not merely personalized but contextually accurate, referencing real internal projects, specific colleagues, and organizational structures scraped from public sources and synthesized at scale. Detection rates for these campaigns using conventional email security tools are reported to be significantly lower than for traditional phishing.
At the network and application layer, AI-driven vulnerability scanning tools — many of them openly available — have compressed the window between CVE disclosure and exploitation from weeks to hours. For security operations centers managing thousands of endpoints, the volume of alerts requiring triage has increased beyond what human analysts can process. The global cybersecurity workforce gap, estimated by ISC2 at approximately 4 million unfilled positions worldwide, amplifies the problem: there are not enough people, and the ones who exist cannot keep up.
CrowdStrike, SentinelOne, and Palo Alto Networks have each made substantial investments in AI-assisted threat detection and autonomous response capabilities, and all three reported record enterprise deal sizes in their most recent fiscal quarters. But the more significant shift is happening among newer entrants.
Companies including Horizon3.ai, Pentera, and Protect AI raised a combined $620 million in 2025 and early 2026 on the premise that continuous automated attack simulation and AI-driven posture management represent the next generation of enterprise defense. Their argument — that organizations need to probe their own systems the way attackers do, at machine speed — is gaining traction with CISOs who have watched traditional penetration testing cycles fail to keep pace with environment changes.
CISA updated its guidance on AI-assisted threat detection in February 2026, explicitly endorsing the use of AI models in SOC environments for tier-one alert triage and anomaly detection, while cautioning organizations to maintain human oversight on response actions with significant operational impact.
The shift is not without tension. Security budgets are not growing fast enough to fund both legacy platform maintenance and AI-native replacements simultaneously. Analysts at Forrester estimate that mid-market enterprises are managing an average of 47 distinct security tools, many with overlapping coverage and fragmented visibility.
The consolidation argument — that fewer, deeper AI-native platforms provide better outcomes than a larger number of point solutions — is resonating in procurement conversations. Several of the largest enterprise deals closed in Q1 2026 involved multi-year platform commitments that explicitly replaced existing tool sets rather than extending them.
The market for AI-native security platforms is projected by IDC to reach $22 billion by 2028. The companies best positioned are those that can demonstrate not just detection capability, but the kind of autonomous response logic that keeps pace with threats that no human analyst can realistically monitor at volume.
]]>The driver is AIP: Palantir’s Artificial Intelligence Platform, which connects enterprise data systems to large language model reasoning in ways that the company claims, with increasing empirical backing, deliver measurable operational outcomes.
What separates Palantir’s enterprise AI motion from competitors is its AIP Bootcamp program — an intensive three-to-five day on-site engagement where Palantir engineers work directly with customer teams to build working AI workflows against the customer’s actual data. The program compresses what would typically be a six-month proof-of-concept into days, and it has proven extraordinarily effective at converting skeptical procurement committees into signed contracts.
Palantir disclosed that it has now run over 680 AIP Bootcamps globally, up from approximately 200 at the start of 2025. The close rate from bootcamp to paid deployment has remained above 70% throughout that expansion, according to comments made by CEO Alex Karp on the earnings call.
The industries leading adoption tell a coherent story: manufacturing (predictive maintenance and supply chain optimization), healthcare (clinical operations and prior authorization automation), and financial services (fraud detection and regulatory reporting). These are not experimental use cases — they are core operational workflows where measurable efficiency gains translate directly to bottom-line impact.
Palantir’s stock has been a lightning rod for debate between those who view its valuation as irrational and those who argue the market is finally pricing in a durable software moat. Q1 2026 gives the bulls fresh material.
US commercial revenue of $275 million in the quarter represents 58% year-over-year growth. Customer count reached 382 US commercial customers, up from 221 a year earlier. Net dollar retention — the measure of how much existing customers expand their spending — stood at 124%, meaning the installed base is growing even before new customer acquisition is counted.
Total company revenue came in at $884 million for the quarter, beating analyst consensus of $851 million. The company raised full-year 2026 guidance to $3.75 billion, implying continued acceleration through the back half of the year.
Palantir’s commercial success raises a question the company has not fully answered: how defensible is AIP against hyperscaler competition? Microsoft, Google, and Amazon are all aggressively building enterprise AI platforms with the advantage of existing deep customer relationships and bundling leverage.
Palantir’s answer, articulated repeatedly by Karp, is that the company’s value is not in the AI models themselves but in the ontology layer — a structured representation of an organization’s data, people, and processes that sits beneath the AI and makes it operable in complex real-world environments. That layer, Palantir argues, takes years to build correctly and cannot be replicated by a cloud provider offering general-purpose tooling.
Whether that moat proves durable at enterprise scale is the defining question for Palantir’s next chapter. The Q1 numbers suggest it is holding, at least for now.
Sources: Palantir Q1 2026 Earnings Release, Palantir AIP Bootcamp Program, Bloomberg Intelligence Enterprise AI Spend Tracker Q1 2026
]]>The timing matters. Nvidia’s Blackwell Ultra and AMD’s MI400 series both require HBM4 at volumes that no single supplier can currently meet. Samsung’s entry into commercial production shifts that calculus.
HBM4 delivers approximately 40% greater memory bandwidth compared to HBM3e, the current generation shipping in production AI accelerators. For large language model inference — where memory bandwidth is frequently the binding constraint on token throughput — this is not an incremental improvement. It is an architectural step change.
Samsung’s HBM4 stacks 16 dies at 1.2TB/s per package, up from HBM3e’s 819GB/s. The company has invested roughly $7.8 billion in dedicated HBM manufacturing lines over the past 18 months, according to capital expenditure filings reviewed by analysts at TF International Securities. That investment is now beginning to yield output at scale.
SK Hynix, which has supplied the bulk of HBM for Nvidia’s H100 and H200 series, held approximately 62% of the global HBM market in 2025 according to Trendforce data. Samsung held around 28%, with Micron making up the remainder. Samsung’s HBM4 ramp is designed to close that gap materially by Q3 2026.
The significance extends beyond market share competition. The AI infrastructure buildout underway globally — from US hyperscalers to sovereign AI programs in the Middle East, Europe, and Southeast Asia — is constrained by HBM availability at least as much as it is by compute silicon. A second viable HBM4 supplier arriving at scale is structurally significant for data center timelines.
Nvidia’s supply agreements with memory manufacturers are not public, but industry analysts widely expect Blackwell Ultra shipments to accelerate in the second half of 2026 if Samsung can qualify at volume. The qualification process — which requires Samsung’s HBM4 to pass rigorous testing inside Nvidia’s package — is reportedly in its final stages, with results expected within six to eight weeks.
AMD, which has historically maintained more diversified memory supply relationships than Nvidia, is expected to begin incorporating Samsung HBM4 into MI400 production variants in Q4 2026.
Samsung’s broader challenge is not just production volume but yield rate. HBM manufacturing requires near-perfect stack bonding across 16 memory dies — a process where defects compound geometrically. SK Hynix’s yield advantage has historically been significant. Whether Samsung has closed that gap at HBM4 will determine whether this mass production announcement translates into a durable competitive repositioning or merely a partial catch-up.
Micron, the third player, is targeting HBM4 sampling in mid-2026 and mass production by year-end, which would further expand supply. For AI infrastructure buyers watching delivery lead times stretch into 2027, more competition in HBM cannot arrive soon enough.
Sources: Trendforce HBM Market Share Report Q1 2026, TF International Securities Samsung CapEx Analysis, Samsung Electronics Q1 2026 Earnings Call Transcript
]]>The company’s pitch is direct: AI agents that investigate incidents, reason across logs, metrics, traces, and change history, and coordinate remediation without waiting for a human to open a ticket.
Production environments have always been the hardest part of software operations. When something breaks at 2 a.m. — a database query starts timing out, a payment API begins returning errors, a deployment silently corrupts a cache layer — the traditional response is a pager alert, a groggy engineer, and an hour of manual log-trawling before anyone even forms a hypothesis.
Resolve AI replaces that sequence with autonomous agents. The platform ingests signals from across the observability stack — logs, metrics, distributed traces, deployment events, configuration changes — and reasons across them in real time. When it identifies an incident, it doesn’t just surface an alert; it builds a causal chain, surfaces the most likely root cause, and in many cases initiates remediation automatically.
Customers including Coinbase, DoorDash, MongoDB, MSCI, Salesforce, and Zscaler are already running the platform in production. For companies at that scale, the economics are clear: the cost of an hour of downtime vastly exceeds the cost of the software that prevents it.
The speed of the valuation step-up is itself a data point. Resolve AI crossed the unicorn threshold with a $125 million Series A announced in February 2026, valuing the company at $1 billion. The April extension — $40 million more at $1.5 billion — signals that investors saw enough early traction to move before a full Series B process.
DST Global’s involvement is particularly notable. The firm, known for late-stage growth bets on companies with demonstrated revenue velocity, typically enters rounds after product-market fit is well established. Salesforce Ventures adds a strategic dimension: Salesforce’s own AI agent platform, Agentforce, operates adjacent to the same enterprise engineering buyer. Whether that adjacency becomes a partnership or an acquisition thesis is an open question.
Alongside the funding, Resolve AI announced the launch of Resolve AI Labs, an internal research unit focused on advancing AI systems for complex production environments — a signal that the company intends to build proprietary model capabilities rather than simply wrapping third-party frontier models.
The broader market context matters. For years, AIOps — the application of machine learning to IT operations — promised more than it delivered. Early tools were good at dashboards and anomaly detection but stopped short of autonomous action. The current generation of large language model-based agents has changed that equation.
Resolve AI is not alone in this space. PagerDuty, ServiceNow, and a handful of well-funded startups are all competing for the same buyer. But Resolve’s focus on causal reasoning in production environments, rather than general IT workflow automation, gives it a defensible technical position. The question for 2026 is whether the platform can scale beyond the mid-market into the largest global enterprises — and whether its autonomous remediation capabilities hold up when the stakes are highest.
At $1.5 billion and growing fast, the market has already formed an opinion.
]]>The raw numbers are staggering. Claude Mythos 5 is built on a refined Mixture of Experts (MoE) architecture, meaning that while the total parameter count sits at 10 trillion, only an estimated 800 billion to 1.2 trillion parameters are active per forward pass. In practical terms, the model carries the knowledge capacity of a 10-trillion-parameter dense system while keeping inference costs closer to a 1-trillion-parameter model.
For context, OpenAI’s GPT-5 and Google’s Gemini Ultra — the previous generation of frontier models — are believed to operate in the 1–2 trillion active parameter range. Mythos 5 represents a leap that independent researchers are describing as the most significant scaling milestone since GPT-4.
Anthropic has confirmed the model’s existence but has not released official benchmarks, a system card, or made it publicly available. The company cited the need for “efficiency improvements and responsible rollout” — language that takes on particular weight given what the leaked documents revealed about the model’s capabilities.
The most striking — and alarming — aspect of Mythos 5 is its performance on cybersecurity tasks. According to Anthropic’s own draft documentation, the model is capable of identifying and exploiting zero-day vulnerabilities across every major operating system and every major web browser when directed by a user to do so. The vulnerabilities it surfaces are often subtle, and the oldest confirmed discovery so far was a now-patched 27-year-old bug in OpenBSD.
This has prompted a controlled early-access program focused specifically on defensive cybersecurity. Select enterprise customers — primarily in critical infrastructure and financial services — are testing the model under strict conditions. Anthropic’s red team has been running structured evaluations to determine the boundaries of what the system can and cannot be permitted to do at general release.
The situation puts Anthropic in an uncomfortable but increasingly familiar position: the company that built the industry’s most articulate case for responsible AI development now holds the most powerful, and potentially most dangerous, model in existence.
The competitive pressure is immediate. OpenAI, Google DeepMind, and Meta’s FAIR lab are all understood to have large-scale MoE projects in advanced development, but none has publicly crossed the 10-trillion mark. If Mythos 5’s capabilities hold up under independent evaluation, it will reset the benchmark bar across reasoning, coding, scientific research, and now offensive security tasks.
For enterprises, the calculus is straightforward: whoever gets early access to Mythos 5 for defensive security applications gains a meaningful advantage in threat detection. For regulators, the story is more complicated. The EU AI Act’s framework for high-risk AI systems was not designed with models capable of finding 27-year-old zero-days in mind — and that gap is about to become very visible.
Anthropic says a broader release timeline will be announced once safety evaluations are complete. Given what’s already leaked, the industry will be watching closely.
]]>Nearly half of all U.S. data centers planned for 2026 — approximately 7 gigawatts of the 12 GW announced — have already been canceled or delayed. Only about one-third of announced 2026 capacity is currently under active construction. The reason is not a shortage of ambition or capital. It is a shortage of power transformers.
Power transformers are the unglamorous infrastructure bottleneck of the AI era. Every data center requires large, custom-engineered transformers to step utility voltage down to the levels that server racks can use. These are not commodity items. Lead times that stretched 24 to 30 months before 2020 have now extended to as long as five years for high-capacity units. AI deployment cycles run under 18 months. The math does not work.
The mismatch is structural. Transformer manufacturing is an industrial process that cannot be scaled quickly. It requires specialized steel, engineered copper windings, long-curing insulation processes, and skilled trades that are not easily automated. No hyperscaler can fix this with a software update or a new chip architecture.
U.S. domestic production covers only approximately 20% of the large power transformers the country needs. The remainder is imported — predominantly from China, which controls roughly 60% of global supply. The two largest Chinese suppliers, TBEA and China XD Group, are reportedly booked through at least 2027.
The situation was already constrained before April 2026. The new U.S. tariff regime has made it considerably worse.
Copper, the primary conductor material in transformer windings, now carries a 50% tariff under the April 2026 duties. Unlike semiconductors — which received targeted exemptions intended to protect the chip supply chain — power equipment and its raw material inputs received no such carveout. The tariffs that were designed in part to reduce dependence on Chinese manufacturing are, in the short term, raising the cost of the only supply chain capable of meeting U.S. data center demand.
The irony is precise: America’s effort to decouple from China in strategic industries is temporarily increasing the cost of the Chinese-made equipment that American data centers depend on, while domestic manufacturing capacity remains years from filling the gap.
The scale of the slowdown is significant even by the standards of an industry accustomed to delays. Of the roughly 12 GW of U.S. data center capacity announced for 2026, the canceled and delayed projects skew heavily toward the speculative tier — capacity announced to attract hyperscaler commitments or demonstrate site readiness, but not yet tied to firm construction contracts.
The projects that are actively building tend to be those with committed anchor tenants, existing utility interconnection agreements, and transformer orders placed well in advance. The companies that moved early on infrastructure procurement — placing transformer orders in 2023 and 2024 — are insulated. Those that waited for lease commitments before ordering equipment are now facing years-long queues.
What the transformer shortage reveals is a specific class of vulnerability that is distinct from the semiconductor debate. Chips are complex, miniaturized, and require specialized fabs to produce. The political and industrial case for domestic chip manufacturing is well-established and has attracted significant federal support through the CHIPS Act.
Large power transformers are not miniaturized. They are heavy industrial products that the United States manufactured domestically for most of the twentieth century. The erosion of that capacity was a function of cost optimization over decades — and it has left the AI infrastructure buildout exposed to a supply chain risk that is simultaneously simpler and harder to fix than the chip shortage.
Simpler, because the technology is not exotic. Harder, because rebuilding industrial manufacturing capacity takes longer than building a fab.
The Inflation Reduction Act has directed some investment toward domestic energy infrastructure, and the transformer shortage has drawn attention from the Department of Energy. But the timeline for meaningful domestic capacity expansion is measured in years — not months — and the data centers that need power interconnections in 2026 and 2027 are running out of room to wait.
The immediate consequence is that the AI infrastructure spending numbers cited by hyperscalers — Microsoft’s $80 billion, Amazon’s $105 billion, Google’s $75 billion — will not translate into online capacity at the pace the press releases imply. Some of that capital will be deployed. Much of it will be queued against infrastructure that doesn’t yet exist.
The medium-term risk is competitive. China’s AI infrastructure buildout does not face the same transformer bottleneck; domestic supply is integrated into the buildout in ways that U.S. programs cannot replicate quickly. If the U.S. data center pipeline remains constrained through 2027 and 2028, the compute capacity gap between the two countries narrows in ways that benchmark comparisons don’t capture.
Big Tech can keep announcing. The transformers will arrive when they arrive.
]]>The company crossed $25 billion in annualized revenue at the end of February 2026, according to reporting by The Information. That’s up from $21.4 billion at year-end 2025 and roughly $6 billion in late 2024 — a trajectory that took roughly 39 months from near-zero. For comparison, Google needed more than five years to reach $25 billion in annual revenue after its 2004 IPO. Salesforce took a decade.
The revenue story is real. The cost story is equally real. OpenAI is projected to lose $14 billion in 2026, according to internal financial projections reviewed by The Information. The company’s cash burn is expected to reach $57 billion annually by 2027, with breakeven not anticipated until 2030. OpenAI has made public commitments to spend $600 billion over five years on infrastructure, data centers, and compute.
That financial picture has created a visible fault line at the executive level. CEO Sam Altman has privately told associates he wants to take OpenAI public as soon as Q4 2026. CFO Sarah Friar disagrees. She has raised concerns about what she describes as the pace of spending commitments and the difficulty of presenting a credible path to profitability to public market investors. According to reports, Altman has since excluded Friar from key financial planning meetings, and she no longer reports to him directly.
OpenAI’s current private valuation stands at approximately $850 billion, following a $110 billion funding round in February 2026 — the largest private technology financing in history. Altman has indicated an IPO target of $1 trillion.
Whether public markets will accept that valuation given the losses depends on how investors frame the story. The optimistic read: OpenAI is building foundational AI infrastructure that will generate massive returns at scale, and early losses are an investment, not a warning sign. The skeptical read: a company burning $57 billion a year by 2027 needs extraordinary revenue growth just to stay solvent, and the competitive pressure from Anthropic, Google DeepMind, and open-source models is intensifying.
Anthropic, for its part, is approaching $19 billion in annualized revenue following its $30 billion Series G raise in March 2026 at a $380 billion valuation — a gap that has narrowed considerably from twelve months ago.
If OpenAI does list in late 2026 or 2027, it would be one of the largest technology IPOs in history, rivaling only SpaceX — which filed its own confidential S-1 this month. The back-to-back potential listings of the two most-watched private companies in tech represent a significant test of how public markets price AI infrastructure in an era of enormous capital deployment and uncertain near-term profitability.
The internal split between Altman and Friar may resolve itself, or it may not. What it signals clearly is that even inside the company widely regarded as setting the pace of the AI race, the financial math of doing so is contested territory.
Sources: The Information — OpenAI tops $25B annualized revenue · WinBuzzer — OpenAI CEO CFO IPO split · The Information — OpenAI losses tripling $14B
]]>