Daily News Coverage: AI, Infrastructure & Tech Business News

The CODEW Daily News Coverage | Tuesday, August 18, 2026

Ten Stories Shaping Today's Technology, Infrastructure, and Capital Markets

AI Inference Prices Keep Falling as Model Competition Shifts to Economics

Enterprise AI costs hit their lowest point of 2026 this month, as OpenAI, Anthropic, and Chinese open-weight labs compete on price as aggressively as capability.

Average enterprise AI inference prices fell to $1.16–$1.18 per million tokens between August 6–8, the lowest level recorded all year, according to Jefferies research citing data from Silicon Data's pricing index, which tracks per-token costs across API providers and open-weight inference platforms. That marks a 43% decline from $2.04 on May 31, and a continued drop from $1.45 in late July — a compression happening faster than most analysts predicted even six months ago.

The decline is being driven from two directions at once. OpenAI cut pricing for its GPT-5.6 model series by up to 80% last month, a move Jefferies frames as evidence of "increasing emphasis on cost efficiencies" across the U.S. AI ecosystem. On the capability side, Jefferies notes that Anthropic's Claude Opus 5 now delivers performance close to its flagship Fable 5 model at roughly half the price — meaning the price war isn't only happening at the low end of the market, but increasingly at the frontier too. Chinese open-weight labs, led by DeepSeek, continue pushing the affordability frontier further, offering competitive inference at a fraction of proprietary API costs and forcing every major lab to defend its pricing position.

Key numbers/companies: Inference prices down 43% since May 31 ($2.04 → $1.16–$1.18 per million tokens) · GPT-5.6 pricing cut up to 80% · Claude Opus 5 priced at roughly half of Fable 5 for comparable performance · OpenAI, Anthropic, DeepSeek, Google all active in the price competition.

Why it matters: Falling inference costs directly expand which enterprise AI workflows are economically viable. A customer-support automation project or a document-processing pipeline that didn't clear an internal ROI threshold in May may clear it easily today, meaning the price war is arguably doing more to accelerate enterprise AI adoption right now than any single model release. It also signals that the industry has largely moved past the phase where raw capability alone justified premium pricing — cost-per-capability, not capability alone, is now the primary lever labs are pulling to win enterprise accounts.

Market implication: The squeeze falls hardest on mid-tier model providers without hyperscaler-scale distribution or balance sheets to absorb sustained price competition. Providers that can't match frontier-adjacent pricing risk losing enterprise accounts to labs willing to compete on cost even at thinner margins, potentially accelerating consolidation among smaller inference providers over the next several quarters.

Enterprise buyers now have genuine leverage to renegotiate AI vendor contracts that were priced under May's cost structure, and the trend shows no sign of reversing as open-weight competition intensifies from Chinese labs.

Sources: South China Morning Post · Jefferies research (via SCMP) · Silicon Data pricing index · OpenAI pricing announcements.

Anthropic, OpenAI and the Race to Turn AI Agents Into Enterprise Products

New OpenAI research shows enterprise AI usage shifting from chat to delegated agent work — with frontier-adopting companies pulling meaningfully ahead of the rest.

OpenAI published new enterprise research this month, titled "Enterprise Signals," showing that as of June 2026, its Codex coding-agent product generated 64% of combined Codex-and-ChatGPT output tokens among enterprise customers — a sharp reversal for a company whose enterprise business was built almost entirely on conversational chat just two years ago. OpenAI describes the shift explicitly as moving "from assistance to delegation," arguing that the companies furthest along in adoption are pulling meaningfully ahead of the rest.

The adoption curve is no longer confined to engineering teams. OpenAI's data shows weekly active enterprise Codex users up 108x in legal, 41x in sales and recruiting, and 26x in marketing since February — compared with a 5x increase in engineering over the same period, suggesting agentic tools are diffusing fastest precisely in the functions that had the least prior exposure to AI coding tools. A companion working paper, "How Organizations Use AI: Evidence from ChatGPT," found ChatGPT Enterprise output tokens grew roughly 7x between June 2025 and March 2026, driven by both new firm adoption and a fourfold usage increase among existing customers.

Anthropic is competing on a parallel track, emphasizing enterprise trust and governance rather than raw usage-growth statistics; Menlo Ventures' market model puts Anthropic at roughly 40% of enterprise LLM spend, up from just 12% in 2023, ahead of OpenAI at 27% and Google at 21% — evidence the two companies are winning enterprise accounts through different playbooks even as both push deeper into agentic products.

Key numbers/companies: Codex = 64% of OpenAI's enterprise output tokens · 108x growth in legal Codex usage since February · ChatGPT Enterprise tokens up ~7x June 2025–March 2026 · Anthropic ~40% of enterprise LLM spend vs. OpenAI 27%, Google 21%.

Why it matters: The shift from conversational assistance to task delegation is the clearest evidence yet that the AI competition enterprises actually care about has moved past which chatbot answers questions best and toward which platform can reliably execute multi-step work with minimal supervision. That reframes the competitive battle entirely — it's no longer just about model benchmarks, but about which company builds the most trustworthy agent execution layer.

Market implication: Enterprise software vendors and system integrators that can help companies operationalize agentic delegation — not just provide model access — stand to capture disproportionate value as this shift accelerates through the rest of 2026.

Sources: OpenAI Enterprise Signals · OpenAI: From Assistance to Execution · BankInfoSecurity · Menlo Ventures market research.

AI Infrastructure Spending Keeps Reshaping the Semiconductor and Data Center Markets

Texas's grid moratorium and CoreWeave's expanding power targets show physical infrastructure, not chip supply, is now the binding constraint on AI growth.

Texas has frozen new data center connections to its power grid pending an audit of energy and water usage across the entire interconnection queue, which has grown to 474 gigawatts of requested capacity — over five times the state's record peak electricity demand. Actual new generation synchronized to the grid over the same period totaled only about 23 gigawatts, less than 11% of the queue's current size, according to ERCOT data. The pause follows a similar move by New York in July and puts an estimated 20% of the entire U.S. data center pipeline at risk of delay, per Bloomberg NEF, with potential revenue losses reaching $8 billion by the first quarter of 2027.

The strain isn't confined to Texas. CoreWeave, one of the fastest-growing specialized AI cloud providers, is targeting at least 8 gigawatts of active power capacity by 2030, up from 1.5 gigawatts today, and just raised its 2026 capex guidance to $35–39 billion specifically to keep pace with contracted demand that is outrunning its own build-out speed. Combined, the five largest hyperscalers remain on pace for roughly $725 billion in 2026 AI capex, up 77% year-over-year — a spending level that increasingly runs into hard physical limits rather than capital constraints.

Key numbers/companies: 474 GW Texas interconnection queue vs. 23 GW connected · $8B potential revenue loss by Q1 2027 (Bloomberg NEF) · CoreWeave targeting 8 GW by 2030 · $725B combined 2026 hyperscaler capex.

Why it matters: For the first time in this AI infrastructure cycle, grid interconnection — not chip availability or capital — is the clearest binding constraint on growth in the country's largest data center markets. That's a fundamentally different bottleneck than the industry has faced before, and one that can't be solved with more capital alone; it requires physical grid buildout and regulatory approval timelines measured in years, not quarters.

Market implication: Operators capable of bringing their own power generation — gas, nuclear, or battery storage — gain a genuine deployment-speed advantage over those depending solely on standard utility interconnection, a variable increasingly factored directly into site-selection and capital-allocation decisions.

Sources: ERCOT interconnection queue data · Bloomberg NEF data center pipeline analysis · CoreWeave Q2 2026 earnings call · Texas Governor's office directive.

Chipmakers Shift From AI Accelerators Toward the Broader AI Hardware Supply Chain

Custom silicon, networking chips, and specialized processors are redistributing value away from merchant GPUs alone.

Every major hyperscaler now runs a maturing custom silicon program that competes directly with merchant GPU purchases for a growing share of AI compute budgets. Microsoft's Maia 200 chip, built on a 3nm process with more than 140 billion transistors, is already deployed in production data centers powering Microsoft 365 Copilot and OpenAI inference workloads. Amazon's Trainium3, its first 3nm chip, is used by both Anthropic and OpenAI — a notable vote of confidence from labs that aren't required to use it. Google's TPU line remains the most mature of the group, and Meta has disclosed four new MTIA chip generations for deployment through 2027.

The most striking recent development is Anthropic reportedly co-designing custom AI inference chips with Samsung serving as manufacturing partner — a move explicitly aimed at reducing dependence on Nvidia GPUs for inference workloads. That puts a pure AI lab, rather than a hyperscaler with decades of chip-design infrastructure, into the custom silicon race directly, suggesting even well-capitalized labs now see owning inference silicon as a competitive necessity rather than a nice-to-have.

Behind nearly all of this custom silicon activity sit Broadcom and Marvell, which together control roughly 95% of the AI ASIC co-design market. Broadcom's AI semiconductor revenue hit $8.4 billion last quarter, up 106% year-over-year, positioning both companies to benefit from custom silicon adoption regardless of which specific hyperscaler chip ultimately wins the most market share.

Key numbers/companies: Broadcom + Marvell control ~95% of AI ASIC co-design · Broadcom AI semiconductor revenue $8.4B, +106% YoY · Microsoft Maia 200, Amazon Trainium3, Google TPU, Meta MTIA all in active deployment · Anthropic/Samsung custom inference chip program.

Why it matters and market implications: Custom silicon is redistributing competitive leverage away from Nvidia's merchant GPU dominance toward hyperscalers and, increasingly, individual AI labs — even as Nvidia's absolute revenue keeps growing on the strength of overall AI demand. For the semiconductor supply chain, the practical effect is that value is spreading across a wider set of co-design and packaging partners rather than concentrating in a single dominant accelerator vendor.

Sources: Broadcom and Marvell earnings disclosures · Microsoft, Amazon, Google, Meta custom silicon disclosures · Industry reporting on Anthropic-Samsung chip partnership.

Neocloud Expansion Puts Pressure on Traditional Cloud Infrastructure Economics

CoreWeave's contracted backlog is growing faster than its own capex guidance, validating the specialist AI cloud model even as it runs on heavy leverage.

CoreWeave reported Q2 2026 revenue of $2.58 billion, up 112% year-over-year, and raised its full-year capital expenditure guidance to $35–39 billion, up from a prior $30–35 billion range, citing accelerating capacity deployment and new customer commitments including Anthropic and Meta. The company's contracted revenue backlog surged to $104 billion, up 246% year-over-year — growth that is outpacing even CoreWeave's own upwardly revised spending plans. Shares jumped roughly 13.5–14% following the report.

The company remains unprofitable, however, posting a Q2 net loss of $626 million, wider than the $290 million loss a year earlier, driven by heavy infrastructure spending and $640 million in quarterly interest expense on $35 billion in balance-sheet debt used to finance GPU purchases — a debt-to-equity ratio of roughly 7.39x. Active power capacity grew to 1.5 gigawatts, with management targeting at least 8 gigawatts by 2030.

That contrast is instructive against Meta's own recent capex guidance raise, which triggered a stock selloff rather than a rally, because Meta lacks a comparable contracted-backlog figure investors can point to as justification for the spending. CoreWeave's backlog-to-capex ratio moving in its favor — backlog growing roughly 15-20 percentage points faster than the capex increase itself — is precisely the signal the market currently rewards.

Key numbers/companies: Q2 revenue $2.58B, +112% YoY · Backlog $104B, +246% YoY · Net loss $626M · Debt-to-equity 7.39x · Capex guidance raised to $35–39B.

Why it matters and market implications: Neoclouds are proving that GPU-focused, contract-driven infrastructure economics can outcompete traditional hyperscaler cloud growth rates, but the model's heavy reliance on debt financing rather than operating cash flow makes it structurally more fragile if AI demand growth decelerates even modestly. Traditional cloud providers are watching closely as CoreWeave's backlog-driven model puts pressure on how AWS, Azure, and Google Cloud justify their own infrastructure spending to investors.

Sources: CoreWeave Q2 2026 earnings call and shareholder letter · CNBC · Meta Q2 2026 earnings comparison.

Identity Theft Becomes a Bigger Ransomware Risk as Attackers Target Credentials

New industry data shows compromised credentials now cause more ransomware attacks than software vulnerabilities, and MFA alone isn't stopping the breaches.

Sophos's State of Ransomware 2026 report, based on a survey of 2,158 IT and cybersecurity leaders across 17 countries whose organizations were hit by ransomware in the past year, found that phishing and malicious email now account for roughly half of all ransomware root causes, with compromised credentials initiating 79% of attacks overall. Vulnerability exploitation, the dominant cause for three consecutive years, fell to 18% from 32% three years earlier — the first time identity-based attacks have overtaken software exploits as ransomware's leading entry point in four years.

The most striking finding concerns multi-factor authentication: MFA was deployed in 97% of compromised-credential ransomware cases, yet still failed to prevent the breach. Attackers are increasingly bypassing MFA through session-token theft, adversary-in-the-middle phishing kits, and "MFA fatigue" attacks that flood a user's device with approval requests until they accidentally tap "approve." Separately, Palo Alto Networks' Unit 42 Global Incident Response Report, drawing on more than 750 engagements, found identity weaknesses played a material role in almost 90% of investigations, with initial access driven by identity-based techniques in 65% of cases.

Unit 42 also found that in 87% of cases, attacker activity crossed multiple attack surfaces — endpoints, identity systems, networks, and cloud services within the same intrusion — meaning defenders increasingly need visibility across an entire estate rather than any single system to catch an attack in progress.

Key numbers/companies: 79% of ransomware attacks start with stolen credentials (Sophos) · MFA deployed in 97% of compromised-credential cases but still failed · ~90% of Unit 42 investigations involved identity weaknesses · 65% of cases had identity-based initial access.

Why it matters and market implication: A decade of security guidance centered on patching software faster no longer matches the dominant attack pattern organizations actually face. Security budgets are likely to reallocate meaningfully toward identity threat detection and response (ITDR), session monitoring, and credential hygiene over the coming quarters — a shift that benefits vendors specializing in identity security over those focused primarily on traditional vulnerability management.

Sources: Dark Reading · Sophos State of Ransomware 2026 · Help Net Security · Palo Alto Networks Unit 42 Global Incident Response Report 2026.

Enterprise Software Vendors Push AI Agents Deeper Into Core Business Workflows

Salesforce and ServiceNow are converting agentic AI into measurable revenue, while Gartner projects a sharp rise in embedded agents across enterprise applications this year.

Salesforce's Agentforce platform has reached roughly $800 million in annual recurring revenue, up 169% year-over-year, in one of the clearest proof points yet that AI agents can generate direct, measurable enterprise software revenue rather than simply serving as a marketing feature layered onto existing products. ServiceNow's Autonomous Workforce agents are separately resolving IT service-desk cases at speeds early customers describe as dramatically faster than human agents handling comparable tickets, with the company targeting autonomous resolution of the large majority of routine IT tickets.

Gartner projects that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from under 5% in 2025 — one of the fastest capability-adoption curves the research firm has tracked in enterprise software. The shift is happening alongside a broader restructuring of how these agents get connected to enterprise data and tools: the Model Context Protocol has reached more than 10,000 enterprise servers and 97 million SDK downloads, while the newer Agent-to-Agent (A2A) protocol, supporting peer-to-peer delegation between agents, is already in production at more than 150 organizations.

Not every deployment is succeeding at the same pace, however. Separate research finds that while 80% of enterprise apps now embed an agent in some form, only 31% actually run one in production, and 88% of pilots never ship — underscoring that embedding the technology and deploying it reliably remain two very different challenges.

Key numbers/companies: Agentforce ARR ~$800M, +169% YoY · Gartner: 40% of apps embedding agents by end-2026 (from <5 10="" 2025="" 88="" 97m="" agent="" downloads="" enterprise="" in="" mcp:="" never="" of="" p="" pilots="" sdk="" servers="" ship.="">

Why it matters and market implication: Enterprise software incumbents with existing workflow ownership and customer relationships are converting AI into revenue meaningfully faster than standalone AI startups without that distribution — a pattern that increasingly favors platform players like Salesforce and ServiceNow over point-solution competitors, even as the underlying adoption gap between embedded and production-deployed agents remains wide.

Sources: Salesforce Agentforce earnings disclosures · ServiceNow Autonomous Workforce product materials · Gartner enterprise AI agent forecasts.

AI Infrastructure Consolidation Accelerates Across Chips, Networking and Data Centers

Strategic buyers are paying premium multiples for scarce physical infrastructure capability, not just AI software or model access.

Eaton's acquisition of liquid-cooling specialist Boyd Corporation for $9.5 billion — against roughly $1.5 billion in Boyd's projected 2026 revenue — is the clearest recent example of infrastructure M&A commanding valuations once reserved almost exclusively for AI software companies. Accenture separately assembled a $4.175 billion OT cybersecurity platform through a majority stake in Dragos and full acquisitions of runZero and NetRise, converting a services-led security position into a software platform ahead of expected critical-infrastructure regulatory mandates.

Elsewhere, S&P Global acquired data-center intelligence firm datacenterHawk to extend its 451 Research technology-intelligence unit, positioning the company to own the benchmarking and analytics layer investors use to track where AI infrastructure capacity is actually being built. Teledyne Technologies' $1.1 billion purchase of imaging-hardware maker Varex Imaging sent Varex shares up roughly 49.5% on announcement — one of the largest single-day M&A premiums of the year, reflecting how scarce sensing and imaging component capacity has become even outside the core AI-chip narrative.

Key numbers/companies: Eaton/Boyd $9.5B · Accenture OT platform (Dragos, runZero, NetRise) $4.175B · Teledyne/Varex $1.1B (Varex +49.5%) · S&P Global/datacenterHawk (undisclosed terms).

Why it matters, and market implications: Buyers across very different sectors — industrials, consulting, financial data, and imaging hardware — are converging on the same conclusion: physical infrastructure capability tied to the AI buildout is scarce enough to justify premium acquisition multiples, independent of whether the target has any direct AI software or model exposure. Expect continued bolt-on consolidation in power, cooling, and physical AI infrastructure categories through year-end as buyers compete for a limited pool of qualified targets.

Sources: Eaton, Accenture, S&P Global, and Teledyne deal announcements and investor materials.

Investors Continue Concentrating Capital on High-Growth AI Infrastructure and Applications

OpenAI and Anthropic alone absorbed 43% of global venture capital last quarter, even as fresh mega-rounds keep flowing to AI-adjacent physical infrastructure startups.

OpenAI and Anthropic together absorbed roughly $217 billion — 43% of all global venture capital deployed — in the most recent quarter, a level of capital concentration with no clear precedent in venture history. That concentration at the very top is reshaping fundraising dynamics across the rest of the AI ecosystem: rather than broad-based AI enthusiasm lifting all categories evenly, most companies outside the frontier-lab tier are raising smaller, sharper rounds with tighter proof requirements.

Even so, megarounds continue flowing to AI-adjacent physical infrastructure and application companies. Enterprise AI platform Fireworks AI raised $1.5 billion at a $17.5 billion valuation. Defense-manufacturing startup Hadrian raised $1.37 billion. Physical-AI and robotics startup Atoms, led by Travis Kalanick, raised $1.7 billion. The common thread across these rounds is a shift toward capital-intensive, physical categories — manufacturing, robotics, energy — that would have been considered atypical venture bets just a few years ago, now being funded at venture scale specifically because they intersect with AI infrastructure demand.

Key numbers/companies: $217B / 43% of global VC to OpenAI + Anthropic combined · Fireworks AI $1.5B at $17.5B valuation · Hadrian $1.37B · Atoms $1.7B.

Why it matters, and market implications: This level of concentration means the private AI financing system's overall stability increasingly rests on a small number of companies continuing to compound at their current trajectory. A slowdown at either frontier lab would ripple outward across venture markets far beyond AI specifically, given how much aggregate capital is now tied to their continued growth. For founders outside the top tier, the practical effect is a bifurcated fundraising environment — extraordinary capital availability at the very top, and much more disciplined, proof-driven terms everywhere else.

Sources: Venture funding disclosures and industry funding trackers · Company funding announcements (Fireworks AI, Hadrian, Atoms).

Analog and Power Electronics Become Strategic Beneficiaries of the AI Infrastructure Buildout

Texas Instruments, Infineon, and their peers are seeing data-center revenue accelerate sharply as AI server racks demand far more power delivery than prior generations.

Texas Instruments' data center revenue grew roughly 90% year-over-year in Q1 2026, accelerating from 70% growth the prior quarter and 50%-plus growth in the quarter before that — a pace significant enough that TI has now broken out data center as its own reporting segment rather than folding it into broader industrial results. CEO Haviv Ilan has pointed to TI's 300mm analog fab advantage, including a new facility in Sherman, Texas, as giving the company a roughly 40% cost advantage over competitors still running older 200mm wafer lines.

Infineon expects revenue from AI data-center power solutions to grow from €1.5 billion in fiscal 2026 to €2.5 billion in fiscal 2027, and has committed an additional €500 million to accelerate capacity expansion. Around 20 global chipmakers, including Infineon, Texas Instruments, and STMicroelectronics, raised power-semiconductor prices 10–25% on July 1, with some high-end power components for AI servers having already surged as much as 85% in a prior pricing round — driven, according to Gartner analyst Sheng Linghai, largely by Nvidia's newer Blackwell-generation chips requiring substantially more power delivery than previous data center hardware.

Analog Devices CEO Vincent Roche has described power delivery as "the vascular system" and power control as "the brain" of AI data centers — a framing that captures why analog and power semiconductor companies, long valued at cyclical industrial multiples, are increasingly being re-rated by analysts as structural AI infrastructure beneficiaries rather than peripheral suppliers.

Key numbers/companies: TI data center revenue +90% YoY (Q1 2026) · Infineon AI power revenue €1.5B → €2.5B (FY26 → FY27) · Power semiconductor price hikes 10-25% (July 1), up to 85% in prior round · Broader power semiconductor market ~$59.9B in 2026 · Texas Instruments, Analog Devices, Infineon, onsemi, Renesas, STMicroelectronics all benefiting.

Why it matters and market implications: As AI server racks scale past 130-140 kilowatts of power density, the "boring" analog and power layer has become a genuine infrastructure bottleneck rather than a commodity afterthought — and pricing power is accumulating in this layer even as industry inventory remains historically low relative to AI server demand, a dynamic expected to persist into Q4 2026 and beyond.

Sources: Texas Instruments and Infineon earnings disclosures · UBS analyst research (via Investing.com) · BigGo Finance reporting on power semiconductor price hikes.