Skip to main content
Compass · Lens Report

State of AI Infrastructure Pressure

A measured view of how technology-press attention is distributed, moving and connected across 40 related topics — who holds attention, who is gaining or cooling, whether attention is concentrating or spreading out, which topics the press repeatedly connects, and what kind of coverage defines the conversation.

This is a Lens Report covering AI Infrastructure Pressure — a measurement of how the technology press covered the 40 topics in this lens, between Jun 2025 and Sep 2026, across 5,952 articles. Measured as of 20 Sep 2026. Coverage draws on at least 89 distinct outlets; the busiest single outlet accounts for about 5.2% of the set's measured press items.

Scope: This lens deliberately looks beneath the chip designers and hyperscaler brands: chips have their own lens (AI Chips & Inference Economics), so the roster here is the build-out layer — data centers, power, grid operators, and the companies running the compute.

New here? How to read this report

Compass reads a curated panel of technology outlets — the same panel, read at the same steady pace, month after month — and counts and classifies what the press publishes about the topics in this lens: how much attention each drew, what kinds of stories ran, which way coverage is moving, and what was covered together. Every figure below is a measurement of that coverage — never advice — and each one opens to the dated articles behind it. Read it as a coverage-landscape briefing — a measured map of the press conversation, not a ranking of the companies or technologies underneath it. The pattern is the product; the conclusions stay yours.

The read in 60 seconds
  • Most-covered topicAlphabet — 13.9% of the lens's coverage
  • Fastest-gainingByteDance — +6.3 pp (6.1% → 12.4% of lens coverage)
  • Attention concentrationThe top 3 topics hold 33% of the coverage — broadly spread
  • Strongest pairingCore Scientific + CoreWeave — covered together in 32 articles · distinctiveness 0.66
What not to over-read: this measures press coverage — not adoption, revenue, or quality.
The measured read

Across AI Infrastructure Pressure’s full history from June 2025 through September 2026, press attention is broadly spread but shifting toward data centers and a wider cast of infrastructure actors as Ollama and named infrastructure threads cool.

Full history 40 topics 5,952 articles
  • Alphabet leads coverage at 13.9%, but ByteDance at 11% and data centers at 8.1% keep the top tier plural, which is why no single subject defines the press agenda.
  • Recent movement compares the latest quarter with the prior quarter: ByteDance gained 6.3 percentage points and data centers gained 3.3 points, so the shift is occurring inside the top tier.
  • Below the leaders, Nscale gained 2.5 percentage points, Nebius 1.9 and Crusoe 1.2, so the expanding edge is distributed across several subjects.
  • Ollama fell 5.9 percentage points, Stargate 1.8 and Colossus 1 1.7, so cooling cuts across both a highly covered name and narrower infrastructure threads.
  • As the press describes progress, 73% of coverage is at announcement, 17% at launch and 7% at results, so the narrative is weighted toward declared activity rather than reported outcomes.
  • Neutral framing accounts for 92%, versus 5% promotional and 2% critical, so shifts in attention occur inside a largely even-toned press narrative.
What Compass measured

This report describes how the technology press covered the 40 topics in this lens over full history — which drew the most coverage, which are gaining or cooling, what's covered together, and how concentrated attention is. It measures coverage patterns, not the products themselves.

Coverage leaders

The coverage set comprises 5,952 articles across 40 subjects, giving the press agenda substantial breadth beyond its leaders. Attention in this lens is broadly spread—the three most-covered topics hold about 33% of the set’s coverage, spread across roughly 18 effective topics. News and updates account for 41%, launches 21% and partnerships 18%, so attention is distributed through recurring event formats as well as across subjects.

TopicTypeShare of coverageMovement
Alphabet company 13.9%+0.5 pp
ByteDance company 11%+6.3 pp
data centers technology 8.1%+3.3 pp
Ollama company 6.6%−5.9 pp
CoreWeave company 5.6%+0.6 pp
Cerebras company 3.6%−0.5 pp
Supermicro company 3.3%+0.8 pp
vLLM technology 3.3%−0.6 pp
Stargate company 3.1%−1.8 pp
Nebius company 2.5%+1.9 pp
Groq company 2.3%+0.9 pp
Nscale company 2.3%+2.5 pp
Corning company 2.3%−0.3 pp
Alibaba Cloud organization 2.2%+1 pp
PJM organization 2%+0.6 pp
Crusoe company 1.9%+1.2 pp

Topic share = that topic's measured articles divided by the sum of every lens topic's measured articles. An article covering several lens topics counts once for each of them, so shares measure attention to topics — not exclusive slices of unique articles. Lens entries are coverage topics, not mutually exclusive market participants — a company, its products and its models can each be topics, and their coverage overlaps. Movement is the share shift described under Momentum shifts below.

Momentum shifts

ByteDance rose from 6.1% to 12.4%, a gain of 6.3 percentage points, making it the largest recent increase and shifting weight inside the established top tier. Data centers climbed from 7.4% to 10.7%, gaining 3.3 points, so capacity-centered coverage is strengthening alongside that company-led move. Ollama’s 5.9-point decline is the clearest cooling move, while Stargate and Colossus 1 also receded, which is why the pattern is a rotation of attention rather than a uniform rise across the set.

Movement is each topic's share of the lens's coverage in the last 3 months versus the 3 months before that, in percentage points (pp). Topics shifting less than 0.2 pp are steady.

Gaining coverage
  • ByteDance+6.3 pp (6.1% → 12.4% of lens coverage)
  • data centers+3.3 pp (7.4% → 10.7% of lens coverage)
  • Nscale+2.5 pp (1.4% → 3.9% of lens coverage)
  • Nebius+1.9 pp (2% → 3.9% of lens coverage)
  • Crusoe+1.2 pp (1.7% → 2.9% of lens coverage)
  • Vera Rubin NVL72+1.1 pp (0.5% → 1.6% of lens coverage)
  • Alibaba Cloud+1 pp (1.6% → 2.6% of lens coverage)
  • Groq+0.9 pp (1.8% → 2.7% of lens coverage)
Cooling
  • Ollama−5.9 pp (10.3% → 4.4% of lens coverage)
  • Stargate−1.8 pp (2.9% → 1.1% of lens coverage)
  • Colossus 1−1.7 pp (2% → 0.3% of lens coverage)
  • IREN−1.6 pp (2.4% → 0.8% of lens coverage)
  • Core Scientific−1.4 pp (2% → 0.6% of lens coverage)
  • Colossus−1.2 pp (2.3% → 1.1% of lens coverage)
  • Chevron−1.2 pp (1.7% → 0.5% of lens coverage)
  • Colossus 2−1.2 pp (1.6% → 0.4% of lens coverage)
Coverage mix over time

How the lens's coverage is divided among its topics, month by month — and how concentrated attention is across the set.

② Share of coverage over time
Each topic's share of this set's topic mentions, period by period (monthly). Absolute counts rise for everyone in a hot category, so share is what reveals whose grip is tightening or loosening. Each column sums to 100%, with the rest of the set folded into Others. One article can name several members of a set — a product family and one of its models, say — so it counts once toward each: read the bands one at a time rather than adding them together.
25%50%75%Jun 25Sep 25Dec 25Mar 26Jun 26Sep 26
Top-3 concentration latest month 33.1%

Steady — across the last three measured months the top three have held 36.8% of this set's coverage, within 3 points of the 35.1% they held at the start of the window. Neither consolidating nor broadening.

Every value here is a dated measurement of coverage within this set, drawn from the Enginerds corpus — the same panel of sources, read at the same steady pace, month after month. It reflects how much the press writes about each topic, not revenue, adoption, or market share. The conclusion stays yours.

What the coverage is made of

Pooled across the whole lens: the kinds of stories the press runs, how far the reported work has progressed (as the press itself describes it), and the press framing.

By story type
News & updates41%
Launch21%
Partnership18%
Funding6%
Regulation4%
Benchmark4%
Everything else6%
By claim stage (as reported)
Announcement73%
Launch17%
Results7%
Speculation3%
By press framing
Neutral92%
Promotional5%
Critical2%
Skeptical1%

Story type is a share of the 4,709 classified coverage items classified on that axis; claim stage of the same number; press framing of the 4,561 whose framing could be read — vendor and primary-source items carry none.

Covered together

Across the full record, the most strongly connected pairing is Core Scientific and CoreWeave, covered together in 32 articles across acquisition, benchmark and partnership reporting, which helps explain why actor coverage groups around CoreWeave. In that same span, Ollama and vLLM share 28 articles, creating a separate coverage strand from the CoreWeave-centered cluster. Data centers and PJM share eight articles and sit in a cluster with ERCOT, so capacity coverage extends into regulation-focused reporting.

Pairings are ranked by connection strength: distinctiveness weighted by how many confirmed events link the pair and how far apart the two topics sit in the field, with a company's own product-family pairs set aside. So a pair can rank above another that shows a higher raw distinctiveness score or more shared articles — each card carries both numbers. Distinctiveness is normalized co-occurrence (NPMI, −1 to 1): how much more often the pair shares stories than the two topics' separate coverage volumes alone would predict. Big topics co-occur often by volume alone; a high score means the pairing itself is the pattern. Each pairing also shows how many outlets its cited stories span: a pairing cited almost entirely by one specialist outlet is that outlet's editorial pattern, not necessarily a field-wide one. "Covered together" is a measurement of shared coverage across Compass's full measured record — not only this report's window — and not a partnership, endorsement, or equivalence.

Coverage clusters

Groups of topics the press tends to cover together — the shape of the domain's coverage neighborhoods.

CoreWeave cluster
CoreWeaveTeraWulfCore ScientificCrusoeIRENStargateFluidstackStarcloudSB EnergyGB300 NVL72Nebius
Core Scientific (32)IREN (14)GB300 NVL72 (7)TeraWulf (7)Fluidstack (4)CoreWeave

Line weight = articles covering the pair together (full measured record).

data centers cluster
data centersPJMERCOTFERCChevronNREL
PJM (8)ERCOT (6)NREL (5)Chevron (4)FERC (3)data centers

Line weight = articles covering the pair together (full measured record).

What to watch

At the gaining edge, Nscale, Nebius and Crusoe added 2.5 percentage points, 1.9 points and 1.2 points, so rising attention is broadening beyond the established leaders. Vera Rubin NVL72 added 1.1 points, while Crusoe and Starcloud share a seven-article launch-and-partnership thread, so peripheral attention is forming around both a named topic and lower-volume actor ties.

Topics in this lens · 40

The full set this report measures, ranked by share of the lens's coverage. Topic share = that topic's measured articles divided by the sum of every lens topic's measured articles. An article covering several lens topics counts once for each of them, so shares measure attention to topics — not exclusive slices of unique articles. Lens entries are coverage topics, not mutually exclusive market participants — a company, its products and its models can each be topics, and their coverage overlaps. Share is never market share, adoption, or importance.

Alphabet company
ByteDance company
data centers technology
Ollama company
CoreWeave company
Cerebras company
Supermicro company
vLLM technology
Stargate company
Nebius company
Groq company
Nscale company
Show all 40 topics
Corning company
Alibaba Cloud organization
PJM organization
Crusoe company
DDN company
ERCOT organization
PCIe 6.0 standard
WEKA company
IREN company
Open Compute Project organization
NREL organization
TeraWulf company
Core Scientific company
Colossus product
Starcloud company
Equinix company
Chevron company
AEMO organization
Digital Realty company
Fluidstack company
Vera Rubin NVL72 product
Colossus 2 product
GB300 NVL72 product
SB Energy company
Hyperion product
Colossus 1 product
Tennessee Valley Authority organization
FERC organization
Representative cited sources
Show all 32 cited sources
Briefing

Press attention is broadly spread but rotating toward data centers and a wider actor roster because Ollama and named infrastructure threads are giving up recent coverage share.

Methodology

Every figure is drawn from the Enginerds corpus — a curated panel of technology outlets, each read at the same steady pace over time — and is a measurement of coverage within this lens: article counts, each topic's share of the set's coverage over time, the direction of that coverage, and which topics are covered together. Titles are shown as published; inclusion is not endorsement.

Topic share = that topic's measured articles divided by the sum of every lens topic's measured articles. An article covering several lens topics counts once for each of them, so shares measure attention to topics — not exclusive slices of unique articles. Lens entries are coverage topics, not mutually exclusive market participants — a company, its products and its models can each be topics, and their coverage overlaps.

Movement is each topic's share of the lens's coverage in the last 3 months versus the 3 months before that, in percentage points (pp). Topics shifting less than 0.2 pp are steady.

Pairings are ranked by connection strength: distinctiveness weighted by how many confirmed events link the pair and how far apart the two topics sit in the field, with a company's own product-family pairs set aside. So a pair can rank above another that shows a higher raw distinctiveness score or more shared articles — each card carries both numbers. Distinctiveness is normalized co-occurrence (NPMI, −1 to 1): how much more often the pair shares stories than the two topics' separate coverage volumes alone would predict. Big topics co-occur often by volume alone; a high score means the pairing itself is the pattern.

What this report does not claim. Compass measures coverage patterns in the technology press. It does not predict outcomes, rank technologies by merit or importance, or recommend what to buy, build, or invest in. Coverage is not the same as real-world significance.

Track this lens live in Compass.

This is a sample of the State of X report Compass generates over any lens or topic set. Explore the lens interactively, or generate your own reports with a Pro plan.

Built on a curated technology press corpus

Compass measures a dated, source-linked corpus of technology coverage — a curated panel of 140+ outlets, specialist publications, and primary sources, each read at the same steady pace, month after month. Every measurement opens back to the cited articles behind it.

View the corpus methodology →