The Tech Atlas

Start at the buyer's side of the table, because that is where the leverage moved. In October 2025 OpenAI, which has no chip factory, announced two supply deals in a week. AMD agreed to supply 6 gigawatts of GPUs and gave OpenAI a warrant for up to 160 million of its own shares, vesting as the purchases scale (AMD). A week later OpenAI and Broadcom said they would deploy 10 gigawatts of accelerators that OpenAI had designed itself (OpenAI).

Nvidia's own offer had come first, a letter of intent for 10 gigawatts and up to $100 billion of investment. Months later Nvidia's finance chief said “We still haven't completed a definitive agreement,” and Jensen Huang called the pledge “never a commitment” (Fortune, Fortune). Put the three side by side and a buyer's strategy shows. A customer that can threaten to design its own chip gets suppliers to bid with equity. AMD welcomed a very large order and handed over a share of itself to win it, and the risk that came in with the order is a customer that now owns part of its supplier.

Look down the rest of the chart for the bottlenecks. TSMC makes 72.5% of the top ten foundries' revenue. SK hynix holds about half of high-bandwidth memory, though Samsung's share is rising as its next generation ships. ASML is the only company that makes the extreme ultraviolet machines every leading-edge fab needs. Where the chart shows one name at a chokepoint, that company sets terms for everyone above it.

The challengers worth watching change the premise of the row. Cerebras builds one chip from an entire wafer, and went public in May 2026. Substrate is trying to replace the light source in lithography with a particle accelerator that writes with X-rays, and it has money and no revenue. In China, Huawei already ships 28% of AI chips to Nvidia's 34% (IDC, as reported). The orders to watch are the 14A decisions at Intel Foundry in the next two quarters, the first racks of OpenAI's own chips, and the volume ramp of AMD's MI450.

Every part of this sector is a queue. GE Vernova has gas turbine slots reserved into 2030, with 56 gigawatts of reservations on its books. Dell ended its last fiscal year with a $43 billion backlog of AI servers, and IDC says supply, not demand, is what holds server shipments back. Equinix summed up its quarter with the line that “the vast majority of our largest deals were driven by AI workloads” (Equinix). A queue is a good place to study strategy, because the ones in line are deciding who they are willing to pay to move up.

Money is the clearest example. Meta needed a gigawatt-scale campus in Louisiana and agreed a joint venture in which Blue Owl funds hold 80% and Meta 20%. Meta leases the buildings and provided a guarantee for the first sixteen years (Meta). Meta welcomed that capital because it let the campus be built faster and kept most of the cost off its own balance sheet. The risk that rode in with it is a long guarantee that stays with Meta if demand disappoints. The first horse on the Trojan Horses page describes how private credit enters a bank's relationships in much the same way.

Power gets bought the same way, years ahead. Meta signed three nuclear deals that support up to 6.6 gigawatts and prepaid Oklo, with a first phase targeted for 2030. Amazon's deal with X-energy could reach 5 gigawatts by 2039. A buyer that pays up front to get something built has taken on some of the builder's risk, which is the same arrangement as Renesas's deposit to Wolfspeed (that case).

The networking row holds a quieter shift. Ethernet has moved from under a fifth of AI cluster switch sales to about two-thirds, and Celestica, which builds to the cloud buyers' own designs, led the second quarter ahead of Nvidia. The startup to watch is Ayar Labs, which wants to replace copper links between chips with light. In storage, VAST Data's $30 billion valuation beside DDN's $5 billion shows how far investors will pay for the fastest-growing name in a row with no audited share figure.

On a September morning Oracle reported that its remaining performance obligations, the contracted revenue it has yet to deliver, had reached $664 billion. In the same release its free cash flow for the quarter was negative $5 billion (Oracle). A company long known for databases has signed the largest AI contracts in the industry and is paying for the buildings before the revenue arrives. Analysts estimate OpenAI is about half of that backlog, so Oracle has made a very large bet on one customer's ability to keep paying.

Google took the opposite route. Its cloud grew 82% in the June quarter and carries a reported $514 billion backlog, and it sells its own chips and its own models alongside the rental. Amazon still leads on share at 28% and grew AWS 37%, its fastest in eighteen quarters. Microsoft's Azure grew 43% and its commercial backlog is reported at $678 billion. The three are growing at the same time, which is unusual, and it tells you the demand behind them is a shortage of capacity and not a contest for customers.

That shortage built a new tier. CoreWeave rents Nvidia systems to the buyers who cannot build quickly enough, and its customers include Meta, OpenAI and Microsoft, which was about 67% of its 2025 revenue. Nebius holds long contracts from Microsoft and Meta. A landlord for the biggest tenants is a strong business while the shortage lasts, and the interest bills show the cost. CoreWeave paid $640 million of net interest in one quarter.

The practice squad is full of companies selling inference, the serving of finished models to developers. Fireworks AI and Together AI both raised at valuations above $8 billion this summer, and Wasabi's no-exit-fee storage shows a different premise, selling the customer's freedom to leave (that case). The test for the whole sector is conversion. A backlog counts promises, and the quarters ahead show how many of them become revenue.

February 2026 was the month the labs became public-company-sized. SpaceX bought xAI in an all-stock deal that closed on the 2nd. OpenAI's $122 billion round followed, at an $852 billion valuation, with Amazon committing $50 billion and Nvidia and SoftBank $30 billion each (Bloomberg Law). Then Anthropic raised $65 billion at $965 billion in May and told investors its annualized revenue had reached $65 billion by July (CNBC). These valuations are reported, and the point is the ordering underneath them. Menlo Ventures' survey put Anthropic first in enterprise spending at about 40%, ahead of OpenAI at 27% and Google at 21%, while Google's Gemini app passed 950 million monthly users.

Anthropic's strategy is to be the neutral buyer. It can use up to a million Google TPUs, trains on Amazon's Trainium, has signed for up to two gigawatts of AMD's chips and still uses Nvidia. Each supplier welcomes the order, and none of them owns the customer. That position gives it a bargaining advantage the single-supplier labs lack, and it explains why chipmakers and clouds are also its investors.

The companies building on the labs sit in a harder spot. Cursor's coding editor buys models from the labs that also sell coding tools, and Menlo estimated Anthropic at 54% of coding model spending. Perplexity competes with Google Search while depending on others' models. These businesses were welcomed by developers for what they add, and the risk inside them is that their suppliers can become their competitors.

Two names are listed as practice squad that raise a different question. Safe Superintelligence is valued near $32 billion with no product revenue. Thinking Machines has a confirmed $12 billion seed valuation and several unconfirmed reports above it. Open-weight models from DeepSeek were reported to take 24.4% of text requests on one developer marketplace in a September week. The thing to watch is whether the frontier stays closed.

In February the software industry was told, by the market, that it had a problem. Commentators put the selloff at about $285 billion, the fear being that AI agents would do the work that companies paid seats to do. A seat is a license for one human, and an agent is not a human. Vendors answered within months. Salesforce began selling agent actions at about ten cents each, according to commentary, and its Agentforce business reached more than $1.5 billion of annual recurring revenue, up over 240% (reported).

The incumbents' strategy is to own the data and the workflow that agents need to act inside. ServiceNow's subscription revenue grew 24.5% in the June quarter. SAP's cloud backlog grew 27%. Microsoft says it has more than 30 million paid Microsoft 365 Copilot seats, up from 15 million in January (TechCrunch, reported). Each of these companies welcomed the agent because it is a new product line. The risk that came with it is that every agent that finishes a task is one less seat to bill.

The startups sell the outcome and skip the seat. Sierra raised $950 million at a valuation above $15 billion to resolve customer-service issues, and Harvey, Decagon and Glean are doing the same in law, support and search. Meanwhile Databricks closed $5 billion at $190 billion in August, and Palantir is growing at 92.8%, which shows the money is following the platforms where data and models meet.

Private equity is watching the same repricing. Silver Lake is reported in talks to take Workday private at a value above $51 billion, and Thoma Bravo completed Dayforce in February. The question for the second half of the year is whether agent revenue replaces seat revenue fast enough, and Workday's slowing total backlog is the number to keep an eye on.

Security used to be a market of dozens of products and one overworked buyer. In 2026 the buyers are consolidating the sellers. Google closed its Wiz purchase on 11 March at about $29.5 billion after adjustments (Alphabet). Palo Alto Networks completed CyberArk in February to “Secure the AI Era,” as its release put it. ServiceNow agreed to pay $7.75 billion for Armis and CrowdStrike reportedly paid $740 million for an identity startup.

The thinking behind each purchase is the same. A chief security officer wants fewer vendors and one place to answer for a breach, so the vendor that sells the most modules to the same customer wins the renewal. Palo Alto's next-generation security ARR is $9.10 billion, up 63%, and part of that growth was bought. CrowdStrike sells flexible credits that a customer can spend across modules, and Falcon Flex accounts now exceed $2.29 billion of ARR.

Identity is where the new fight is, because AI agents need identities and permissions as people do. Okta says more than 90% of its customers run agents in production and only 22% are confident the agents are controlled, which is a company-reported figure with an obvious interest behind it. CyberArk now sits inside Palo Alto, and CrowdStrike bought SGNL, so Okta is competing against platforms that treat identity as one module among many.

The Wiz case shows a promise that does real work. Google committed to keep the product available on AWS, Azure and Oracle Cloud, and that is what keeps its value as a neutral layer (the note on that deal). The startups to watch are Island, valued at $6.4 billion, and Cyera, reported at $12 billion, both selling a control point that the platforms do not yet own.

Meta's second quarter is a good place to see how ads work now. Its advertising revenue rose 27% to $59.36 billion, on an audience that grew 3% to 3.60 billion daily people. Impressions rose 14% and the price of each ad rose 12%, so it earned far more from the same people by showing them better matches (secondary report). Google's near 90% of search referral traffic and Amazon's $19.8 billion of ads belong to the same story, the three companies sell access to attention at the moment of intent.

Streaming shows what happens when the attention is already crowded. YouTube reached a record 14.2% of US television time in July, the largest single service, while Netflix held 7.9% in June. Netflix's operating margin was 34.6% in its latest quarter and it still cut its ad target to $2.7 billion and about 1,000 jobs (Netflix). On 6 October Paramount closed its $31 a share cash purchase of Warner Bros. Discovery, a scale answer to the same pressure.

TikTok's US venture closed in January, with Oracle, Silver Lake and MGX holding 15% each and ByteDance keeping 19.9%. Daily US use held near 95% of earlier levels. Each new owner welcomed the user base, and the thing to watch is whether the recommendation system stays as good under their control.

The sector's open question is what happens when the buyer is an AI agent. Shopify reported $115.6 billion of volume, up 32%, with AI-driven orders tripling from a small base. Google's AI Mode passed a billion monthly users. Roblox, at the other end, saw bookings slow to 8% as spending per hour fell. The chart's first row shows who earns the ad dollar today. The rows beneath it show who is trying to move where the question gets asked.

The most surprising line on this chart connects a phone to a memory chip. IDC counted 276.3 million smartphones shipped in the second quarter, down 7.4%, and described a memory crisis that is disrupting the market with unprecedented costs and strained supply (IDC, as reported). It expects 2026 shipments to fall 16.7%, the steepest drop on record, as the average price rises 27.6% to $581. The DRAM that AI data centers are buying in bulk is the same DRAM that goes into the phone.

In a shortage the strongest brand gets the supply. Samsung and Apple were the only top-five phone makers to grow, with Samsung up 8.1% and Apple up 14.9%, while Xiaomi, OPPO and vivo fell between 17% and 26% in one IDC-based table. Samsung also makes memory, which gives it a second way to win. In PCs the same split appears. IDC counted 68.2 million shipments, down 4.9%, and Apple was the only major maker to grow, up about 10% with its new MacBook Neo.

The category that is growing is the one nobody had to buy a new phone for. Smart glasses and headsets grew about 35% in the quarter, with glasses at nearly 85% of the category. Meta holds roughly 69% to 76% of smart glasses, depending on the source, and EssilorLuxottica sold more than 7 million of the Meta-branded pairs in 2025. Meta paused the international launch of its Ray-Ban Display glasses because United States demand was backlogged.

Google and Samsung plan Android XR glasses for the fall, with no confirmed date. The row to watch is how long memory stays scarce. IDC expects the shortage to last into early 2028, which means the device makers' margins depend on suppliers in a different sector.

Visa announced that its payments volume passed $4 trillion in a single quarter for the first time, up 10% at constant dollars, with net revenue of $11.6 billion (Investing.com). Mastercard's gross dollar volume was $2.9 trillion. Both sell the same thing, a fee on each transaction on a network that banks and merchants have no practical way to leave, and both are growing their value-added services faster than their core, 34% and 20%. The cards are the foundation, and the services are where the networks are looking for new revenue.

The pressure on them comes from outside the card. Stablecoin supply is about $316 to $322 billion, with Tether near 58% and Circle's USDC at $73.3 billion at the end of the second quarter (Circle). Circle earns the interest on the reserves behind its coins, and the return rate fell 66 basis points to 3.48% in the quarter, so its revenue grew 7% while average circulation grew 25%. An issuer's income moves with interest rates in a way a card network's does not.

The most dramatic move was a proposal. In mid-July Stripe and Advent were reported to have offered $60.50 a share for PayPal, about $53.4 billion and a 28% premium, financed with $17 billion of equity that includes Block and about $50 billion of bank debt (CNBC). Stripe processed $1.9 trillion in 2025, so the younger company would own the older one jointly with a buyout firm. I found no later reporting on PayPal's response, and the cards say so.

Among the banks, Revolut's reported $115 billion valuation from an employee share sale sits next to public markets valuing listed neobanks at roughly four to five times revenue. Chime's public figures, with revenue of $670 million, up 27%, are the cleanest data point in the row, and it counts active members, which makes it hard to compare with Nubank's registered customers.